WO2024043760A1 - Method, device, and recording medium for image encoding/decoding - Google Patents

Abstract

A method, a device, and a recording medium for image encoding/decoding are disclosed. Provided in embodiments are a method, a device and a recording medium which efficiently implement a lightweight in-loop filter and a lightweight predictive deep neural network in order to enhance the encoding efficiency of an image and reduce the time required for encoding. Provided in embodiments is learning in a lightweight in-loop filter and a lightweight predictive deep neural network through a knowledge distillation technique. Provided through embodiments is a learning method in an in-loop filter and a predictive deep neural network for enhancing the encoding efficiency and speed of an image.

Description

Method, device and recording medium for video encoding/decoding

The present invention relates to a method, device, and recording medium for video encoding/decoding.

The present invention claims the benefit of the filing date of Korean Patent Application No. 10-2022-0107154 filed on August 25, 2022 and the benefit of the filing date of Korean Patent Application No. 10-2023-0111834 filed on August 25, 2023, The entire contents are incorporated into this specification.

Through the continued development of the information and communications industry, broadcasting services with HD (High Definition) resolution have spread globally. Through this proliferation, many users have become accustomed to high-resolution, high-definition images and/or videos.

In order to satisfy users' demand for high image quality, many organizations are accelerating the development of next-generation imaging devices. User interest in not only High Definition TV (HDTV) and Full HD (FHD) TV, but also Ultra High Definition (UHD) TV, which has a resolution more than four times that of FHD TV. has increased, and with this increase in interest, image encoding/decoding technology for images with higher resolution and image quality is required.

As video compression technology, there are various technologies such as inter prediction technology, intra prediction technology, transformation and quantization technology, and entropy coding technology.

Inter prediction technology is a technology that predicts the value of a pixel included in the current picture using pictures before and/or after the current picture. Intra prediction technology is a technology that predicts the value of a pixel included in the current picture using information about the pixel in the current picture. Transformation and quantization technology is a technology for compressing the energy of the residual image. Entropy coding technology is a technology that assigns short codes to values with a high frequency of occurrence and long codes to values with a low frequency of occurrence.

Using this video compression technology, video data can be effectively compressed, transmitted, and stored.

One embodiment may provide a method, apparatus, and recording medium that provides learning in an in-loop filter and prediction deep neural network.

In one aspect, selecting a deep neural network; Generating encoded deep neural network information by performing encoding on deep neural network information used to construct the deep neural network; and generating a bitstream including the encoded deep neural network information.

The image encoding method may further include performing learning in the deep neural network.

The step of performing learning in the deep neural network includes performing learning in a teacher network; performing first learning on the student network; And it may include performing second learning in the student network.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

The input to the in-loop filter may be an image in which compression deterioration has occurred.

The output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed so that the error between the output of the in-loop filter and the original image is minimized.

The deep neural network may be a predictive deep neural network that generates a predicted image similar to the original image through prediction.

On the other hand, receiving a bitstream including encoded deep neural network information; Generating deep neural network information used to construct a deep neural network by decoding the encoded deep neural network information; And an image decoding method including the step of selecting the deep neural network is provided.

The image decoding method may further include performing learning in the deep neural network.

The step of performing learning in the deep neural network includes performing learning in a teacher network; performing first learning on the student network; And it may include performing second learning in the student network.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

The input to the in-loop filter may be an image in which compression deterioration has occurred.

The output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed so that the error between the output of the in-loop filter and the original image is minimized.

The deep neural network may be a predictive deep neural network that generates a predicted image similar to the original image through prediction.

On another side, in a computer-readable recording medium storing a bitstream for image decoding, the bitstream includes encoded deep neural network information, and decoding the encoded deep neural network information is performed. Deep neural network information used to construct a deep neural network is generated, and a computer-readable recording medium on which the deep neural network is selected is provided.

Learning in the deep neural network may be performed.

Learning in the teacher network can be performed.

First learning in the student network may be performed.

Secondary learning may be performed in the student network.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

The input to the in-loop filter may be an image in which compression deterioration has occurred.

The output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed so that the error between the output of the in-loop filter and the original image is minimized.

Methods, apparatus, and recording media are provided that provide learning in in-loop filter and prediction deep neural networks.

1 is a block diagram showing the configuration of an encoding device to which the present invention is applied according to an embodiment.

Figure 2 is a block diagram showing the configuration of a decoding device according to an embodiment to which the present invention is applied.

Figure 3 is a diagram schematically showing the division structure of an image when encoding and decoding an image.

Figure 4 is a diagram showing the form of a prediction unit that a coding unit can include.

Figure 5 is a diagram showing the form of a conversion unit that can be included in a coding unit.

Figure 6 shows division of a block according to an example.

Figure 7 is a diagram for explaining an embodiment of the intra prediction process.

Figure 8 is a diagram for explaining reference samples used in the intra prediction process.

Figure 9 is a diagram for explaining an embodiment of the inter prediction process.

Figure 10 shows spatial candidates according to an example.

Figure 11 shows the order of adding motion information of spatial candidates to a merge list according to an example.

Figure 12 explains the process of conversion and quantization according to an example.

13 shows diagonal scanning according to an example.

14 shows horizontal scanning according to an example.

15 shows vertical scanning according to one example.

Figure 16 is a structural diagram of an encoding device according to an embodiment.

Figure 17 is a structural diagram of a decoding device according to an embodiment.

Figure 18 is a flowchart of an encoding method according to an embodiment.

Figure 19 is a flowchart of a decryption method according to an embodiment.

Figure 20 is a flow diagram of learning in a deep neural network according to one example.

Figure 21 shows learning in a teacher network according to an example.

Figure 22 shows learning in an intra-prediction deep neural network and an inter-prediction deep neural network according to one embodiment.

23 shows first learning in a student network according to one embodiment.

24 illustrates first learning in a student network according to one embodiment.

25 illustrates secondary learning in a student network according to one embodiment.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention.

For a detailed description of the exemplary embodiments described below, refer to the accompanying drawings, which illustrate specific embodiments by way of example. These embodiments are described in sufficient detail to enable those skilled in the art to practice the embodiments. It should be understood that the various embodiments are different from one another but are not necessarily mutually exclusive. For example, specific shapes, structures and characteristics described herein with respect to one embodiment may be implemented in other embodiments without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the embodiment. Accordingly, the detailed description that follows is not to be taken in a limiting sense, and the scope of the exemplary embodiments is limited only by the appended claims, together with all equivalents to what those claims assert if properly described.

Similar reference numbers in the drawings refer to identical or similar functions across various aspects. The shapes and sizes of elements in the drawings may be exaggerated for clearer explanation.

In the present invention, terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another. For example, a first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component without departing from the scope of the present invention. The term “and/or” may include any of a plurality of related stated items or a combination of a plurality of related stated items.

When a component is said to be “connected” or “connected” to another component, the two components may be directly connected or connected to each other, but It should be understood that other components may exist in the middle of the components. On the other hand, when a component is said to be “directly connected” or “directly connected” to another component, it should be understood that no other component exists in between the two components. something to do.

Components appearing in the embodiments are shown independently to represent different characteristic functions, and do not mean that each component consists of separate hardware or a single software component. In other words, each component is listed and included as a separate component for convenience of explanation, and at least two of each component are combined to form one component, or one component is divided into multiple components to function. It can be performed, and integrated embodiments and separate embodiments of each of these components are included in the scope of the present invention as long as they do not deviate from the essence of the present invention.

The terms used in the examples are only used to describe specific examples and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In embodiments, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are intended to indicate the presence of one or more other features, numbers, steps, operations, components, parts, or combinations thereof. It should be understood that this does not exclude in advance the presence or addition of features, numbers, steps, operations, components, parts, or combinations thereof. In other words, the description of “including” a specific configuration in the embodiments does not exclude configurations other than the configuration, and means that additional configurations may also be included in the practice of the present invention or the scope of the technical idea of the present invention. .

In embodiments, the term “at least one” may mean one of one or more numbers, such as 1, 2, 3, and 4. In embodiments, the term “a plurality of” may mean one of two or more numbers, such as 2, 3, and 4.

Some of the components of the embodiments may not be essential components that perform essential functions in the present invention, but may simply be optional components to improve performance. Embodiments may be implemented by including only components essential for implementing the essence of the embodiments, excluding components used only to improve performance. Structures that include only essential components excluding optional components used to improve performance are also included in the scope of the embodiments.

Hereinafter, embodiments will be described in detail with reference to the attached drawings so that those skilled in the art can easily implement the embodiments. In describing the embodiments, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present specification, the detailed description will be omitted. In addition, the same reference numerals are used for the same components in the drawings, and duplicate descriptions of the same components are omitted.

Hereinafter, an image may refer to a picture constituting a video, and may also represent the video itself. For example, “encoding and/or decoding of an image” may mean “encoding and/or decoding of a video,” and may mean “encoding and/or decoding of one of the images that constitute a video.” It may be possible.

Hereinafter, the terms “video” and “motion picture(s)” may be used with the same meaning and may be used interchangeably.

Hereinafter, the target image may be an encoding target image that is the target of encoding and/or a decoding target image that is the target of decoding. Additionally, the target image may be an input image input to an encoding device or may be an input image input to a decoding device. Additionally, the target video may be a current video that is currently subject to encoding and/or decoding. For example, the terms “target image” and “current image” may be used interchangeably and may have the same meaning.

Hereinafter, the terms “image”, “picture”, “frame” and “screen” may be used with the same meaning and may be used interchangeably.

Hereinafter, the target block may be an encoding target block that is the target of encoding and/or a decoding target block that is the target of decoding. Additionally, the target block may be a current block that is currently the target of encoding and/or decoding. For example, the terms “target block” and “current block” may be used interchangeably and may be used interchangeably. The current block may mean an encoding target block that is the target of encoding during encoding and/or a decoding target block that is the target of decoding during decoding. Additionally, the current block may be at least one of a coding block, a prediction block, a residual block, and a transform block.

Hereinafter, the terms “block” and “unit” may be used with the same meaning and may be used interchangeably. Alternatively, “block” may refer to a specific unit.

Hereinafter, the terms “region” and “segment” may be used interchangeably.

In embodiments, each of the specified information, data, flag, index, element, attribute, etc. may have a value. The value "0" of information, data, flags, indexes, elements, and attributes may represent false, logical false, or a first predefined value. That is, the values “0”, false, logical false and the first predefined value can be used interchangeably. The value "1" in information, data, flags, indexes, elements, and attributes may represent true, logical true, or a second predefined value. That is, the value “1”, true, logical true and the second predefined value can be used interchangeably.

When a variable such as i or j is used to represent a row, column, or index, the value of i may be an integer greater than or equal to 0, or an integer greater than or equal to 1. That is, in embodiments rows, columns, indices, etc. may be counted from 0, and may be counted from 1.

In embodiments, the term “one or more” or the term “at least one” may mean the term “plural.” “One or more” or “at least one” can be used interchangeably with “plural.”

Below, terms used in the embodiments are explained.

Encoder: An encoder may refer to a device that performs encoding. In other words, an encoder may mean an encoding device.

Decoder: A decoder may refer to a device that performs decoding. In other words, a decoder may mean a decryption device.

Unit: A unit may represent a unit of encoding and/or decoding of an image. The terms “unit” and “block” may be used interchangeably and may be used interchangeably.

- A unit may be an MxN array of samples. M and N can each be positive integers. A unit can often refer to an array of samples in a two-dimensional form.

- In video encoding and decoding, a unit may be an area created by dividing one video. In other words, a unit may be a specified area within one image. One image can be divided into multiple units. Alternatively, a unit may refer to the divided parts when one image is divided into segmented parts and encoding or decoding is performed on the segmented parts.

- In encoding and decoding of images, predefined processing may be performed on the unit depending on the type of unit.

- Depending on the function, the types of units include Macro Unit, Coding Unit (CU), Prediction Unit (PU), Residual Unit, and Transform Unit (TU), etc. It can be classified as: Alternatively, depending on the function, the unit may be a block, macroblock, Coding Tree Unit, Coding Tree Block, Coding Unit, Coding Block, or Prediction Unit. It may mean Prediction Unit, Prediction Block, Residual Unit, Residual Block, Transform Unit, Transform Block, etc. For example, the target unit may be at least one of a CU, PU, residual unit, and TU that are the target of encoding and/or decoding.

- A unit may refer to information including a luma component block, a corresponding chroma component block, and a syntax element for each block in order to refer to it separately from a block.

- The size and shape of the unit may vary. Additionally, units may have various sizes and shapes. In particular, the shape of the unit may include geometric shapes that can be expressed in two dimensions, such as squares, rectangles, trapezoids, triangles, and pentagons.

- Additionally, the unit information may include at least one of the unit type, unit size, unit depth, unit encoding order, and unit decoding order. For example, the type of unit may indicate one of CU, PU, residual unit, and TU.

- One unit can be further divided into subunits having a smaller size than the unit.

Depth: Depth may refer to the degree to which a unit is divided. Additionally, the depth of a unit may indicate the level at which the unit(s) exist when the unit(s) are expressed as a tree structure.

- Unit division information may include depth regarding the depth of the unit. Depth may indicate the number and/or extent to which a unit is divided.

- In a tree structure, the root node has the shallowest depth, and the leaf node has the deepest depth. The root node may be the highest node. A leaf node may be the lowest node.

- One unit can be hierarchically divided into a plurality of sub-units while having depth information based on a tree structure. In other words, a unit and a sub-unit created by division of the unit may respectively correspond to a node and a child node of the node. Each divided sub-unit can have depth. Since depth indicates the number and/or extent to which a unit is divided, division information of a sub-unit may include information about the size of the sub-unit.

- In a tree structure, the highest node may correspond to the first undivided unit. The highest node may be referred to as the root node. Additionally, the highest node may have the minimum depth value. At this time, the highest node may have a depth of level 0.

- A node with a depth of level 1 may represent a unit created as the original unit is divided once. A node with a depth of level 2 may represent a unit created by dividing the original unit twice.

- A node with a depth of level n may represent a unit created as the original unit is divided n times.

- A leaf node may be the lowest node and may be a node that cannot be further divided. The depth of the leaf node may be at the maximum level. For example, the predefined value of the maximum level may be 3.

- QT depth can indicate the depth of quad division. BT depth may indicate the depth for binary division. TT depth can indicate depth to strikeout splits.

Sample: A sample may be the base unit that constitutes a block. Samples can be expressed as values from 0 to 2 ^Bd -1 depending on the bit depth (Bd).

- Samples can be pixels or pixel values.

- Hereinafter, the terms “pixel”, “pixel” and “sample” may be used with the same meaning and may be used interchangeably.

Coding Tree Unit (CTU): A CTU may be composed of one luma component (Y) coding tree block and two chroma component (Cb, Cr) coding tree blocks related to the luma component coding tree block. there is. Additionally, CTU may mean including the above blocks and syntax elements for each block of the above blocks.

- Each coding tree unit is composed of sub-units such as coding unit, prediction unit, and transformation unit, such as Quad Tree (QT), Binary Tree (BT), and Ternary Tree (TT). It can be partitioned using one or more partitioning methods. Quad tree may refer to a quarternary tree. Additionally, each coding tree unit may be partitioned using a MultiType Tree (MTT) using one or more partitioning methods.

- CTU can be used as a term to refer to a pixel block, which is a processing unit in the decoding and encoding process of an image, as in segmentation of an input image.

Coding Tree Block (CTB): Coding tree block can be used as a term to refer to any one of Y coding tree block, Cb coding tree block, and Cr coding tree block.

Neighbor block: A neighboring block may refer to a block adjacent to the target block. A neighboring block may also mean a reconstructed neighboring block.

- Hereinafter, the terms “neighboring block” and “adjacent block” may be used with the same meaning and may be used interchangeably.

- A neighbor block may mean a reconstructed neighbor block.

Spatial neighboring block: A spatial neighboring block may be a block spatially adjacent to the target block. Neighboring blocks may include spatial neighboring blocks.

- The target block and spatial neighboring blocks may be included in the target picture.

- A spatial neighboring block may refer to a block bordering the target block or a block located within a predetermined distance from the target block.

- A spatial neighboring block may refer to a block adjacent to the vertex of the target block. Here, the block adjacent to the vertex of the target block may be a block vertically adjacent to a neighboring block horizontally adjacent to the target block, or a block horizontally adjacent to a neighboring block vertically adjacent to the target block.

Temporal neighboring block: A temporal neighboring block may be a block temporally adjacent to the target block. A neighboring block may include temporal neighboring blocks.

- Temporal neighboring blocks may include co-located blocks (col blocks).

- A call block may be a block in an already reconstructed co-located picture (col picture). The position of the call block within the call picture may correspond to the position of the target block within the target picture. Alternatively, the location of the collocated block in the collocated picture may be the same as the location of the target block in the target picture. A call picture may be a picture included in the reference picture list.

- A temporal neighboring block may be a block temporally adjacent to a spatial neighboring block of the target block.

Prediction mode: The prediction mode may be information indicating the mode used for intra prediction or the mode used for inter prediction.

Prediction unit: A prediction unit may refer to a base unit for prediction such as inter prediction, intra prediction, inter compensation, intra compensation, and motion compensation.

- One prediction unit may be divided into a plurality of partitions or sub-prediction units with smaller sizes. A plurality of partitions may also be a basic unit in performing prediction or compensation. A partition created by dividing a prediction unit may also be a prediction unit.

Prediction unit partition: Prediction unit partition may refer to the form in which the prediction unit is divided.

Reconstructed neighboring unit: The reconstructed neighboring unit may be a unit that has already been decrypted and rebuilt in the neighborhood of the target unit.

- The reconstructed neighboring unit may be a spatially adjacent unit or a temporally adjacent unit to the target unit.

- The reconstructed spatial neighboring unit may be a unit in the target picture and a unit that has already been reconstructed through encoding and/or decoding.

- The reconstructed temporal neighboring unit may be a unit in the reference image and a unit that has already been reconstructed through encoding and/or decoding. The location of the reconstructed temporal neighboring unit in the reference image may be the same as the location of the target unit in the target picture, or may correspond to the location of the target unit in the target picture. Alternatively, the reconstructed temporal neighboring unit may be a neighboring block of the corresponding block in the reference image. Here, the location of the corresponding block within the reference image may correspond to the location of the target block within the target image. Here, that the positions of the blocks correspond may mean that the positions of the blocks are the same, that one block is included in another block, and that one block occupies the specified position of the other block. It could mean doing it.

Sub-picture: A picture can be divided into one or more sub-pictures. A sub-picture may consist of one or more tile rows and one or more tile columns.

- A sub-picture may be an area with a square or rectangular (i.e., non-square) shape within the picture. Additionally, the sub-picture may include one or more CTUs. .

- A sub-picture may be a rectangular area of one or more slices within one picture.

- One sub-picture may include one or more tiles, one or more bricks, and/or one or more slices.

Tiles: A tile can be an area within a picture that has a square or rectangular shape (i.e., a non-square shape).

- A tile may contain one or more CTUs.

- A tile can be split into one or more bricks.

Brick: A brick may refer to one or more CTU rows within a tile.

- A tile can be split into one or more bricks. Each brick may contain one or more CTU rows.

- Tiles that are not divided into two or more can also refer to bricks.

Slice: A slice may contain one or more tiles within a picture. Alternatively, a slice may include one or more bricks within a tile.

- A sub-picture may contain one or more slices that collectively cover a rectangular area within the picture. Accordingly, each sub-picture boundary may always be a slice boundary. Additionally, each vertical sub-picture boundary may always be a vertical tile boundary.

Parameter set: The parameter set may correspond to header information among the structures in the bitstream.

- Parameter sets include Video Parameter Set (VPS), Sequence Parameter Set (SPS), Picture Parameter Set (PPS), Adaptation Parameter Set (APS), and decoding parameters. It may include at least one of a set (Decoding Parameter Set (DPS)), etc.

Information signaled through a parameter set can be applied to pictures referencing the parameter set. For example, information in the VPS can be applied to pictures referencing the VPS. Information in the SPS can be applied to pictures referencing the SPS. Information in the PPS can be applied to pictures referencing the PPS.

A parameter set can refer to a parent parameter set. For example, PPS may refer to SPS. SPS may refer to VPS.

- Additionally, the parameter set may include tile group, slice header information, and tile header information. A tile group may refer to a group including a plurality of tiles. Additionally, the meaning of a tile group may be the same as that of a slice.

Rate-distortion optimization: The coding device uses a combination of coding unit size, prediction mode, prediction unit size, motion information, and transformation unit size to provide high coding efficiency. Distortion optimization can be used.

- The rate-distortion optimization method can calculate the rate-distortion cost of each combination to select the optimal combination among the above combinations. The rate-distortion cost can be calculated using the formula “D+λ*R”. In general, the combination that minimizes the rate-distortion cost according to the formula "D+λ*R" can be selected as the optimal combination in the rate-distortion optimization method.

- D may indicate distortion. D may be the mean square error of the difference values between the original transform coefficients and the reconstructed transform coefficients within the transform unit.

- R can represent a rate. R can represent the bit rate using related context information.

- λ may represent a Lagrangian multiplier. R may include not only coding parameter information such as prediction mode, motion information, and coded block flag, but also bits generated by encoding of transform coefficients.

- The encoding device may perform processes such as inter prediction, intra prediction, transformation, quantization, entropy coding, inverse quantization, and/or inverse transformation to calculate accurate D and R. These processes can greatly increase the complexity of the encoding device.

Bitstream: A bitstream may refer to a string of bits containing encoded video information.

Parsing: Parsing may mean entropy decoding a bitstream to determine the value of a syntax element. Alternatively, parsing may mean entropy decoding itself.

Symbol: May refer to at least one of a syntax element, a coding parameter, and a transform coefficient of an encoding target unit and/or a decoding target unit. Additionally, a symbol may mean the object of entropy encoding or the result of entropy decoding.

Reference picture: A reference picture may refer to an image that a unit refers to for inter prediction or motion compensation. Alternatively, the reference picture may be an image that includes a reference unit referenced by the target unit for inter prediction or motion compensation.

Hereinafter, the terms “reference picture” and “reference image” may be used with the same meaning and may be used interchangeably.

Reference picture list: A reference picture list may be a list containing one or more reference pictures used for inter prediction or motion compensation.

- The types of reference picture lists are List Combined (LC), List 0 (L0), List 1 (L1), List 2 (L2), and List 3 (List 3; L3). ), etc.

- One or more reference picture lists can be used in inter prediction.

Inter prediction indicator: The inter prediction indicator may indicate the direction of inter prediction for the target unit. Inter prediction can be either one-way prediction or two-way prediction. Alternatively, the inter prediction indicator may indicate the number of reference pictures used when generating a prediction unit of the target unit. Alternatively, the inter prediction indicator may mean the number of prediction blocks used for inter prediction or motion compensation for the target unit.

Prediction list utilization flag: The prediction list utilization flag may indicate whether a prediction unit is generated using at least one reference picture in a specific reference picture list.

- An inter prediction indicator can be derived using the prediction list utilization flag. Conversely, the prediction list utilization flag can be derived using the inter prediction indicator. For example, when the prediction list utilization flag indicates the first value of 0, it may indicate that a prediction block is not generated using a reference picture in the reference picture list for the target unit. When the prediction list utilization flag indicates a second value of 1, it may indicate that a prediction unit is generated using a reference picture list for the target unit.

Reference picture index: The reference picture index may be an index that indicates a specific reference picture in the reference picture list.

Picture Order Count (POC): The POC of a picture may indicate the display order of the picture.

Motion Vector (MV): A motion vector may be a two-dimensional vector used in inter prediction or motion compensation. A motion vector may mean an offset between a target image and a reference image.

- For example, MV can be expressed in the form (mv _x , mv _y ). mv _x can represent the horizontal component, and mv _y can represent the vertical component.

Search range: The search range may be a two-dimensional area where a search for MV is performed during inter prediction. For example, the size of the search area may be MxN. M and N can each be positive integers.

Motion vector candidate: A motion vector candidate may refer to a block that is a prediction candidate or a motion vector of a block that is a prediction candidate when predicting a motion vector.

- The motion vector candidate may be included in the motion vector candidate list.

Motion vector candidate list: The motion vector candidate list may refer to a list constructed using one or more motion vector candidates.

Motion vector candidate index: The motion vector candidate index may refer to an indicator indicating a motion vector candidate in the motion vector candidate list. Alternatively, the motion vector candidate index may be the index of a motion vector predictor.

Motion information: Motion information includes motion vectors, reference picture indices, and inter prediction indicators, as well as reference picture list information, reference pictures, motion vector candidates, motion vector candidate indices, merge candidates, and merge indices. It may mean information containing at least one of the following.

Merge candidate list: The merge candidate list may refer to a list constructed using one or more merge candidates.

Merge candidates: Merge candidates include spatial merge candidates, temporal merge candidates, combined merge candidates, combined bi-prediction merge candidates, candidates based on history, candidates based on the average of two candidates, and zero. It may mean a merge candidate, etc. The merge candidate may include an inter prediction indicator and may include motion information such as a reference picture index for each list, a motion vector, a prediction list utilization flag, and an inter prediction indicator.

Merge index: The merge index may be an indicator pointing to a merge candidate in the merge candidate list.

- The merge index may indicate the reconstructed unit that derived the merge candidate among the reconstructed units that are spatially adjacent to the target unit and the reconstructed units that are temporally adjacent to the target unit.

- The merge index may indicate at least one of the motion information of the merge candidate.

Transform unit: A transform unit may be a basic unit in residual signal coding and/or residual signal decoding, such as transform, inverse transform, quantization, inverse quantization, transform coefficient coding, and transform coefficient decoding. One transformation unit may be divided into a plurality of sub-transformation units with smaller sizes. Here, the transformation may include one or more of a first-order transformation and a second-order transformation, and the inverse transformation may include one or more of a first-order inversion and a second-order inversion.

Scaling: Scaling may refer to the process of multiplying the transform coefficient level by a factor.

- As a result of scaling to the transform coefficient level, a transform coefficient can be generated. Scaling may also be referred to as dequantization.

Quantization Parameter (QP): A quantization parameter may refer to a value used when generating a transform coefficient level for a transform coefficient in quantization. Alternatively, the quantization parameter may refer to a value used when generating a transform coefficient by scaling the transform coefficient level in dequantization. Alternatively, the quantization parameter may be a value mapped to the quantization step size.

Delta quantization parameter: The delta quantization parameter may mean the difference between the predicted quantization parameter and the quantization parameter of the target unit.

Scan: Scan can refer to a method of sorting the order of coefficients within a unit, block, or matrix. For example, sorting a two-dimensional array into a one-dimensional array can be called scanning. Alternatively, arranging a one-dimensional array into a two-dimensional array can also be referred to as a scan or inverse scan.

Transform coefficient: The transform coefficient may be a coefficient value generated as the encoding device performs transformation. Alternatively, the transformation coefficient may be a coefficient value generated as the decoding device performs at least one of entropy decoding and inverse quantization.

- Quantized levels or quantized transform coefficient levels generated by applying quantization to the transform coefficient or residual signal may also be included in the meaning of the transform coefficient.

Quantized level: A quantized level may refer to a value generated by performing quantization on a transform coefficient or residual signal in an encoding device. Alternatively, the quantized level may mean a value that is the target of inverse quantization when performing inverse quantization in a decoding device.

- The quantized transform coefficient level, which is the result of transformation and quantization, can also be included in the meaning of the quantized level.

Non-zero transform coefficient: A non-zero transform coefficient may mean a transform coefficient with a non-zero value or a transform coefficient level with a non-zero value. Alternatively, a non-zero transform coefficient may mean a transform coefficient whose value size is not 0 or a transform coefficient level whose value size is not 0.

Quantization matrix: A quantization matrix may refer to a matrix used in a quantization process or dequantization process to improve the subjective or objective image quality of an image. The quantization matrix may also be referred to as a scaling list.

Quantization matrix coefficient: The quantization matrix coefficient may refer to each element in the quantization matrix. Quantization matrix coefficients may also be referred to as matrix coefficients.

Default matrix: The default matrix may be a quantization matrix predefined in the encoding device and the decoding device.

Non-default matrix: A non-default matrix may be a quantization matrix that is not predefined in the encoding device and the decoding device. A non-default matrix may refer to a quantization matrix signaled by a user from an encoding device to a decoding device.

Most Probable Mode (MPM): MPM may indicate an intra prediction mode that is likely to be used for intra prediction of the target block.

- The encoding device and the decoding device may determine one or more MPMs based on coding parameters related to the target block and properties of entities related to the target block.

- The encoding device and the decoding device may determine one or more MPMs based on the intra prediction mode of the reference block. There may be multiple reference blocks. The plurality of reference blocks may include a spatial neighboring block adjacent to the left of the target block and a spatial neighboring block adjacent to the top of the target block. In other words, one or more different MPMs may be determined depending on which intra prediction modes are used for the reference blocks.

- One or more MPMs can be determined in the same way in the encoding device and the decoding device. In other words, the encoding device and the decoding device may share an MPM list containing the same one or more MPMs.

MPM list: The MPM list may be a list containing one or more MPMs. The number of one or more MPMs in the MPM list may be predefined.

MPM indicator: The MPM indicator may indicate the MPM used for intra prediction of the target block among one or more MPMs in the MPM list. For example, the MPM indicator may be an index to the MPM list.

- Since the MPM list is determined in the same way in the encoding device and the decoding device, the MPM list itself may not need to be transmitted from the encoding device to the decoding device.

- The MPM indicator can be signaled from the encoding device to the decoding device. As the MPM indicator is signaled, the decoding device can determine the MPM to be used for intra prediction for the target block among the MPMs in the MPM list.

MPM usage indicator: The MPM usage indicator may indicate whether the MPM usage mode will be used for prediction of the target block. The MPM use mode may be a mode that uses the MPM list to determine the MPM to be used for intra prediction for the target block.

- The MPM use indicator can be signaled from the encoding device to the decoding device.

Signaling: Signaling may indicate that information is transmitted from an encoding device to a decoding device. Alternatively, signaling may mean that an encoding device includes information in a bitstream or recording medium. Information signaled by the encoding device can be used by the decoding device.

- The encoding device can generate encoded information by performing encoding on signaled information. Encoded information can be transmitted from an encoding device to a decoding device. The decoding device can obtain information by performing decoding on the transmitted encoded information. Here, encoding may be entropy encoding, and decoding may be entropy decoding.

Selective signaling: Information can be signaled selectively. Selective signaling of information may mean that an encoding device selectively includes information (according to certain conditions) in a bitstream or recording medium. Selective signaling of information may mean that the decoding device selectively extracts information from the bitstream (according to specific conditions).

Omission of signaling: Signaling of information may be omitted. Omission of signaling about information may mean that the encoding device does not include the information (depending on certain conditions) in the bitstream or recording medium. Omission of signaling for information may mean that the decoding device does not extract information from the bitstream (according to certain conditions).

Statistical value: Variables, coding parameters, constants, etc. may have values that can be operated on. The statistical value may be a value generated by an operation on the values of these specified objects. For example, a statistical value may be the average value, weighted average value, weighted sum, minimum value, maximum value, mode, etc. for values of specified variables, specified coding parameters, and specified constants. Can be one or more of values, intermediate values, and interpolated values.

1 is a block diagram showing the configuration of an encoding device to which the present invention is applied according to an embodiment.

The encoding device 100 may be an encoder, a video encoding device, or an image encoding device. A video may contain one or more images. The encoding device 100 may sequentially encode one or more images of a video.

Referring to FIG. 1, the encoding device 100 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, and an entropy encoding unit. It may include a unit 150, an inverse quantization unit 160, an inverse transform unit 170, an adder 175, a filter unit 180, and a reference picture buffer 190.

The encoding device 100 may perform encoding on the target image using intra mode and/or inter mode. In other words, the prediction mode for the target block may be one of intra mode and inter mode.

Hereinafter, the terms “intra mode”, “intra prediction mode”, “intra-picture mode” and “intra-picture prediction mode” may be used with the same meaning and may be used interchangeably.

Hereinafter, the terms “inter mode”, “inter prediction mode”, “inter-screen mode”, and “inter-screen prediction mode” may be used with the same meaning and may be used interchangeably.

Hereinafter, the term “image” may refer only to a portion of an image or may refer to a block. Additionally, processing of “image” may represent sequential processing of a plurality of blocks.

Additionally, the encoding device 100 can generate a bitstream including encoded information through encoding of a target image, and output and store the generated bitstream. The generated bitstream may be stored in a computer-readable recording medium and streamed through wired and/or wireless transmission media.

As the prediction mode, if intra mode is used, switch 115 can be switched to intra. As the prediction mode, if the inter mode is used, the switch 115 can be switched to inter.

The encoding device 100 may generate a prediction block for the target block. Additionally, after the prediction block is generated, the encoding device 100 may encode the residual block for the target block using the residual of the target block and the prediction block.

When the prediction mode is intra mode, the intra prediction unit 120 may use pixels of a block that is already encoded and/or decoded in the neighborhood of the target block as a reference sample. The intra prediction unit 120 may perform spatial prediction for the target block using a reference sample and generate prediction samples for the target block through spatial prediction. A prediction sample may refer to a sample within a prediction block.

The inter prediction unit 110 may include a motion prediction unit and a motion compensation unit.

When the prediction mode is inter mode, the motion prediction unit can search for the area that best matches the target block from the reference image during the motion prediction process, and use the searched area to derive motion vectors for the target block and the searched area. can do. At this time, the motion prediction unit may use the search area as the area that is the target of the search.

The reference image may be stored in the reference picture buffer 190, and when encoding and/or decoding of the reference image is processed, the encoded and/or decoded reference image may be stored in the reference picture buffer 190.

As the decoded picture is stored, the reference picture buffer 190 may be a decoded picture buffer (DPB).

The motion compensation unit may generate a prediction block for the target block by performing motion compensation using a motion vector. Here, the motion vector may be a two-dimensional vector used for inter prediction. Additionally, the motion vector may indicate an offset between the target image and the reference image.

When the motion prediction unit and the motion compensation unit have a non-integer value, the motion prediction unit and the motion compensation unit may generate a prediction block by applying an interpolation filter to some areas in the reference image. In order to perform inter prediction or motion compensation, the methods of motion prediction and motion compensation of the PU included in the CU based on the CU include skip mode, merge mode, and advanced motion vector prediction. Prediction (AMVP) mode or current picture reference mode can be determined, and inter prediction or motion compensation can be performed according to each mode.

The subtractor 125 may generate a residual block, which is the difference between the target block and the prediction block. A residual block may also be referred to as a residual signal.

The residual signal may refer to the difference between the original signal and the predicted signal. Alternatively, the residual signal may be a signal generated by transforming or quantizing, or transforming and quantizing, the difference between the original signal and the predicted signal. A residual block may be a residual signal on a block basis.

The transform unit 130 may generate a transform coefficient by performing transformation on the residual block and output the generated transform coefficient. Here, the transformation coefficient may be a coefficient value generated by performing transformation on the residual block.

The conversion unit 130 may use one of a plurality of predefined conversion methods when performing conversion.

The plurality of predefined transformation methods may include Discrete Cosine Transform (DCT), Discrete Sine Transform (DST), and Karhunen-Loeve Transform (KLT) based transformation, etc. there is.

The transformation method used to transform the residual block may be determined according to at least one of coding parameters for the target block and/or the neighboring block. For example, the conversion method may be determined based on at least one of the inter prediction mode for the PU, the intra prediction mode for the PU, the size of the TU, and the shape of the TU. Alternatively, conversion information indicating a conversion method may be signaled from the encoding device 100 to the decoding device 200.

When the transform skip mode is applied, the transform unit 130 may omit transforming the residual block.

A quantized transform coefficient level or quantized level can be generated by applying quantization to the transform coefficient. Hereinafter, in embodiments, the quantized transform coefficient level and the quantized level may also be referred to as transform coefficients.

The quantization unit 140 may generate a quantized transform coefficient level (that is, a quantized level or a quantized coefficient) by quantizing the transform coefficient according to the quantization parameter. The quantization unit 140 may output the generated quantized transform coefficient level. At this time, the quantization unit 140 may quantize the transform coefficient using a quantization matrix.

The entropy encoding unit 150 may generate a bitstream by performing entropy encoding according to a probability distribution based on the values calculated by the quantization unit 140 and/or coding parameter values calculated during the encoding process. . The entropy encoding unit 150 may output the generated bitstream.

The entropy encoding unit 150 may perform entropy encoding on information about pixels of an image and information for decoding the image. For example, information for decoding an image may include syntax elements, etc.

When entropy coding is applied, a small number of bits may be assigned to symbols with a high probability of occurrence, and a large number of bits may be assigned to symbols with a low probability of occurrence. As symbols are expressed through this allocation, the size of the bitstring for the symbols that are the target of encoding can be reduced. Therefore, the compression performance of video encoding can be improved through entropy coding.

In addition, the entropy encoding unit 150 uses exponential golomb, context-adaptive variable length coding (CAVLC), and context-adaptive binary arithmetic coding for entropy encoding. Coding methods such as Arithmetic Coding (CABAC) can be used. For example, the entropy encoding unit 150 may perform entropy encoding using a Variable Length Coding/Code (VLC) table. For example, the entropy encoding unit 150 may derive a binarization method for the target symbol. Additionally, the entropy encoder 150 can derive a probability model of the target symbol/bin. The entropy encoding unit 150 may perform arithmetic encoding using the derived binarization method, probability model, and context model.

The entropy encoder 150 can change the coefficients of the two-dimensional block form into the form of a one-dimensional vector through a transform coefficient scanning method to encode the quantized transform coefficient level.

Coding parameters may be information required for encoding and/or decoding. The coding parameter may include information encoded in the encoding device 100 and transmitted from the encoding device 100 to the decoding device, and may include information that can be derived during the encoding or decoding process. For example, information transmitted to the decoding device includes syntax elements.

Coding parameters are encoded in an encoding device, such as syntax elements, and may include information derived from the encoding process or decoding process, as well as information (or flags and indexes, etc.) signaled from the encoding device to the decoding device. there is. Additionally, coding parameters may include information required when encoding or decoding an image. For example, the size of the unit/block, the shape of the unit/block, the depth of the unit/block, the division information of the unit/block, the division structure of the unit/block, information indicating whether the unit/block is divided into a quad tree form, Information indicating whether the unit/block is split into a binary tree, the direction of the binary tree split (horizontal or vertical), the split type of the binary tree (symmetric split or asymmetric split), and whether the unit/block is split into a ternary tree. Information indicating whether the unit/block is divided into a ternary tree, the direction of division (horizontal or vertical), the division type of the ternary tree (symmetric division or asymmetric division, etc.), and whether the unit/block is a multi-type tree. Information indicating whether the division is divided in the form of a multi-type tree, the combination and direction of the division in the multi-type tree form (horizontal or vertical direction, etc.), the division type of the division in the multi-type tree form (symmetric division or asymmetric division), multi-type Splitting tree in the form of a tree (binary tree or ternary tree), type of prediction mode (intra prediction or inter prediction), intra prediction mode/direction, intra luma prediction mode/direction, intra chroma prediction mode/direction, intra splitting information, inter Segmentation information, coding block segmentation flag, prediction block segmentation flag, transformation block segmentation flag, reference sample filtering method, reference sample filter tab (tap), reference sample filter coefficient, prediction block filtering method, prediction block filter tab, prediction block filter coefficient , prediction block boundary filtering method, prediction block boundary filter tab, prediction block boundary filter coefficient, inter prediction mode, motion information, motion vector, motion vector difference, reference picture index, inter prediction direction, inter prediction indicator, prediction list utilization. ) Flag, reference picture list, reference image, POC, motion vector predictor, motion vector prediction index, motion vector prediction candidate, motion vector candidate list, information indicating whether merge mode is used, merge index, merge candidate, merge candidate list , information indicating whether skip mode is used, type of interpolation filter, filter tab of the interpolation filter, filter coefficient of the interpolation filter, motion vector size, motion vector expression accuracy, transformation type, transformation size, and first-order transformation. Information indicating whether to use, information indicating whether to use additional (secondary) transformation, primary transformation selection information (or primary transformation index), secondary transformation selection information (or secondary transformation index), residual Information indicating the presence or absence of a signal, coded block pattern, coded block flag, quantization parameter, residual quantization parameter, quantization matrix, information on intra-loop filter, intra-loop filter Information indicating whether to apply, coefficient of intra-loop filter, filter tab of intra-loop, shape/form of intra-loop filter, information indicating whether to apply deblocking filter, deblocking filter Coefficients, filter tab of deblocking filter, strength of deblocking filter, shape/form of deblocking filter, information indicating whether adaptive sample offset is applied, adaptive sample offset value, adaptive sample offset category, adaptive sample Offset type, information indicating whether to apply an adaptive in-loop filter, coefficients of the adaptive-loop filter, filter tab of the adaptive-loop filter, shape of the adaptive-loop filter/ Shape, binarization/debinarization method, context model, context model determination method, context model update method, information indicating whether regular mode is performed, information indicating whether bypass mode is performed, significant coefficient flag, and finally Significant coefficient flag, coefficient group unit coding flag, last significant coefficient position, flag indicating whether the coefficient value is greater than 1, flag indicating whether the coefficient value is greater than 2, flag indicating whether the coefficient value is greater than 3 , remaining coefficient value information, sign information, reconstructed luma sample, reconstructed chroma sample, context bin, bypass bin, residual luma sample, residual chroma sample, transformation coefficient, luma transformation coefficient, chroma transformation coefficient, Quantized level, luma quantized level, chroma quantized level, transform coefficient level, luma transform coefficient level, chroma transform coefficient level, transform coefficient level scanning method, size of motion vector search area on the side of the decoding device, size of the motion vector search area on the side of the decoding device Shape of motion vector search area on the side, number of motion vector searches on the side of the decoder, CTU size, minimum block size, maximum block size, maximum block depth, minimum block depth, video display/output order, slice identification information , slice type, slice division information, tile group identification information, tile group type, tile group division information, tile identification information, tile type, tile division information, picture type, bit depth, input sample bit depth, reconstructed sample bit depth. , at least one value of residual sample bit depth, transform coefficient bit depth, quantized level bit depth, information about luma signal, information about chroma signal, color space of target block, and color space of residual block, Combined forms or statistics may be included in coding parameters. Additionally, information related to the above-described coding parameters may also be included in the coding parameters. Information used to calculate and/or derive the coding parameters described above may also be included in the coding parameters. Information calculated or derived using the above-described coding parameters may also be included in the coding parameters.

Primary transformation selection information may indicate the primary transformation applied to the target block.

Secondary transformation selection information may indicate secondary transformation applied to the target block.

The residual signal may represent the difference between the original signal and the predicted signal. Alternatively, the residual signal may be a signal generated by transforming the difference between the original signal and the predicted signal. Alternatively, the residual signal may be a signal generated by converting and quantizing the difference between the original signal and the predicted signal. A residual block may be a residual signal for a block.

Here, signaling information may mean that the encoding device 100 includes entropy-encoded information generated by performing entropy encoding on a flag or index in a bitstream. , this may mean that the decoding device 200 obtains information by performing entropy decoding on entropy-encoded information extracted from the bitstream. Here, the information may include flags and indexes.

A signal may refer to signaled information. Hereinafter, information about images and blocks may be referred to as signals. Additionally, hereinafter, the terms “information” and “signal” may be used with the same meaning and may be used interchangeably. For example, a specific signal may be a signal representing a specific block. The original signal may be a signal representing the target block. A prediction signal may be a signal representing a prediction block. The residual signal may be a signal representing a residual block.

The bitstream may include information according to a specified syntax. The encoding device 100 may generate a bitstream including information according to a specified syntax. The encoding device 200 may obtain information from the bitstream according to the specified syntax.

Since encoding through inter prediction is performed by the encoding device 100, the encoded target image can be used as a reference image for other image(s) to be processed later. Accordingly, the encoding device 100 can reconstruct or decode the encoded target image, and store the reconstructed or decoded image as a reference image in the reference picture buffer 190. For decoding, inverse quantization and inverse transformation may be processed on the encoded target image.

The quantized level may be inversely quantized in the inverse quantization unit 160 and inversely transformed in the inverse transformation unit 170. The inverse quantization unit 160 may generate an inverse quantized coefficient by performing inverse quantization on the quantized level. The inverse transform unit 170 may generate inverse quantized and inverse transformed coefficients by performing inverse transformation on the inverse quantized coefficients.

The inverse-quantized and inverse-transformed coefficients can be combined with the prediction block through the adder 175. A reconstructed block can be generated by combining the inverse-quantized and inverse-transformed coefficients with the prediction block. Here, the dequantized and/or inverse-transformed coefficient may mean a coefficient on which at least one of dequantization and inverse-transformation has been performed, and may mean a reconstructed residual block. Here, the reconstructed block may mean a recovered block or a decoded block.

The reconstructed block may pass through the filter unit 180. The filter unit 180 includes at least a deblocking filter, a sample adaptive offset (SAO), an adaptive loop filter (ALF), and a non-local filter (NLF). One or more can be applied to a reconstructed sample, reconstructed block, or reconstructed picture. The filter unit 180 may also be referred to as an in-loop filter.

The deblocking filter can remove block distortion occurring at the boundaries between blocks in the reconstructed picture. To determine whether to apply a deblocking filter, it may be determined whether to apply a deblocking filter to the target block based on the pixel(s) included in a few columns or rows included in the block.

When applying a deblocking filter to a target block, the applied filter may vary depending on the strength of deblocking filtering required. In other words, among different filters, a filter determined according to the strength of deblocking filtering may be applied to the target block. When a deblocking filter is applied to the target block, a long-tap filter, strong filter, weak filter, and Gaussian filter are used depending on the strength of the deblocking filtering required. ) one or more filters may be applied to the target block.

Additionally, when vertical filtering and horizontal filtering are performed on the target block, horizontal filtering and vertical filtering may be processed in parallel.

SAO can add an appropriate offset to the pixel value of a pixel to compensate for coding errors. SAO can perform correction using an offset for the difference between the original image and the deblocked image in pixel units for the image to which deblocking has been applied. In order to perform offset correction on an image, a method is used to divide the pixels included in the image into a certain number of areas, determine the area where offset is to be performed among the divided areas, and apply the offset to the determined area. A method of applying an offset by considering edge information of each pixel of the image may be used.

ALF can perform filtering based on a comparison between the reconstructed image and the original image. After dividing the pixels included in the image into predetermined groups, a filter to be applied to each divided group can be determined, and filtering can be performed differentially for each group. Information related to whether to apply an adaptive loop filter may be signaled for each CU. This information can be signaled for the luma signal. The shape of the ALF and filter coefficients to be applied to each block may be different for each block. Alternatively, regardless of the characteristics of the block, a fixed form of ALF may be applied to the block.

The non-local filter can perform filtering based on reconstructed blocks similar to the target block. An area similar to the target block may be selected from the reconstructed image, and filtering of the target block may be performed using statistical properties of the selected similar area. Information related to whether to apply a non-local filter may be signaled to the CU. Additionally, the shapes and filter coefficients of non-local filters to be applied to blocks may be different depending on the block.

The reconstructed block or reconstructed image that has passed through the filter unit 180 may be stored in the reference picture buffer 190 as a reference picture. The reconstructed block that has passed through the filter unit 180 may be part of a reference picture. In other words, the reference picture may be a reconstructed picture composed of reconstructed blocks that have passed through the filter unit 180. The stored reference picture can then be used for inter prediction or motion compensation.

Figure 2 is a block diagram showing the configuration of a decoding device according to an embodiment to which the present invention is applied.

The decoding device 200 may be a decoder, a video decoding device, or an image decoding device.

Referring to FIG. 2, the decoding device 200 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, and a switch 245. , may include an adder 255, a filter unit 260, and a reference picture buffer 270.

The decoding device 200 may receive the bitstream output from the encoding device 100. The decoding device 200 can receive a bitstream stored in a computer-readable recording medium and can receive a bitstream streaming through a wired/wireless transmission medium.

The decoding device 200 may perform intra-mode and/or inter-mode decoding on the bitstream. Additionally, the decoding device 200 can generate a reconstructed image or a decoded image through decoding, and output the generated reconstructed image or a decoded image.

For example, switching to intra mode or inter mode according to the prediction mode used for decoding may be performed by the switch 245. If the prediction mode used for decoding is intra mode, the switch 245 may be switched to intra mode. If the prediction mode used for decoding is the inter mode, the switch 245 may be switched to inter.

The decoding device 200 can obtain a reconstructed residual block by decoding the input bitstream and generate a prediction block. When the reconstructed residual block and the prediction block are obtained, the decoding device 200 can generate a reconstructed block that is the target of decoding by combining the reconstructed residual block and the prediction block.

The entropy decoding unit 210 may generate symbols by performing entropy decoding on the bitstream based on a probability distribution for the bitstream. The generated symbols may include symbols in the form of quantized transform coefficient levels (i.e., quantized levels or quantized coefficients). Here, the entropy decoding method may be similar to the entropy encoding method described above. For example, the entropy decoding method may be the reverse process of the entropy encoding method described above.

The entropy decoder 210 can change the coefficients in the form of a one-dimensional vector into the form of a two-dimensional block through a transform coefficient scanning method in order to decode the quantized transform coefficient level.

For example, by scanning the coefficients of a block using a diagonal scan in the upper right corner, the coefficients can be changed into a two-dimensional block form. Alternatively, which scan to use among the upper right diagonal scan, vertical scan, and horizontal scan may be determined depending on the block size and/or intra prediction mode.

The quantized coefficient may be inverse quantized in the inverse quantization unit 220. The inverse quantization unit 220 may generate an inverse quantized coefficient by performing inverse quantization on the quantized coefficient. Additionally, the inverse quantized coefficient may be inversely transformed in the inverse transform unit 230. The inverse transform unit 230 may generate a reconstructed residual block by performing inverse transform on the inverse quantized coefficients. As a result of performing inverse quantization and inverse transformation on the quantized coefficients, a reconstructed residual block may be generated. At this time, the inverse quantization unit 220 may apply a quantization matrix to the quantized coefficients when generating a reconstructed residual block.

When the intra mode is used, the intra prediction unit 240 may generate a prediction block by performing spatial prediction on the target block using pixel values of already decoded blocks neighboring the target block.

The inter prediction unit 250 may include a motion compensation unit. Alternatively, the inter prediction unit 250 may be called a motion compensation unit.

When the inter mode is used, the motion compensation unit may generate a prediction block by performing motion compensation on the target block using a motion vector and a reference image stored in the reference picture buffer 270.

When the motion vector has a non-integer value, the motion compensation unit can apply an interpolation filter to some areas in the reference image and generate a prediction block using the reference image to which the interpolation filter has been applied. In order to perform motion compensation, the motion compensation unit may determine which of skip mode, merge mode, AMVP mode, and current picture reference mode is the motion compensation method used for the PU included in the CU based on the CU, and the determined mode. Motion compensation can be performed according to .

The reconstructed residual block and prediction block can be added through an adder 255. The adder 255 may generate a reconstructed block by adding the reconstructed residual block and the prediction block.

The reconstructed block may pass through the filter unit 260. The filter unit 260 may apply at least one of a deblocking filter, SAO, ALF, and non-local filter to the reconstructed block or the reconstructed image. The reconstructed image may be a picture containing reconstructed blocks.

The filter unit 260 may output a reconstructed image.

The reconstructed block and/or the reconstructed image that has passed through the filter unit 260 may be stored as a reference picture in the reference picture buffer 270. The reconstructed block that has passed through the filter unit 260 may be part of a reference picture. In other words, the reference picture may be a reconstructed image composed of reconstructed blocks that have passed through the filter unit 260. The stored reference picture can then be used for inter prediction and/or motion compensation.

Figure 3 is a diagram schematically showing the division structure of an image when encoding and decoding an image.

Figure 3 may schematically show an example in which one unit is divided into a plurality of sub-units.

In order to efficiently segment an image, a coding unit (CU) may be used in encoding and decoding. A unit may be a term that refers to a combination of 1) a block containing video samples and 2) a syntax element. For example, “division of a unit” may mean “division of a block corresponding to a unit.”

CU may be used as a base unit for video encoding and/or decoding. Additionally, a CU may be used as a unit to which a selected mode of intra mode and inter mode is applied in video encoding and/or decoding. In other words, in video encoding and/or decoding, it can be determined which mode among intra mode and inter mode will be applied to each CU.

Additionally, a CU may be a basic unit in prediction, transformation, quantization, inverse transformation, inverse quantization, and encoding and/or decoding of transformation coefficients.

Referring to FIG. 3, the image 300 may be sequentially divided into units of largest coding units (LCUs). For each LCU, a partition structure may be determined. Here, LCU may be used with the same meaning as Coding Tree Unit (CTU).

Division of a unit may mean division of a block corresponding to the unit. Block division information may include depth information regarding the depth of the unit. Depth information may indicate the number and/or extent to which a unit is divided. One unit may be hierarchically divided into a plurality of sub-units with depth information based on a tree structure.

Each divided sub-unit may have depth information. Depth information may be information indicating the size of the CU. Depth information may be stored for each CU.

Each CU may have depth information. When a CU is split, CUs created by splitting may have a depth that increases by 1 from the depth of the split CU.

The division structure may refer to the distribution of CUs within the LCU 310 for efficiently encoding images. This distribution may be determined depending on whether to divide one CU into multiple CUs. The number of divided CUs may be a positive integer greater than or equal to 2, including 2, 4, 8, and 16.

The horizontal and vertical sizes of the CU created by division may be smaller than the horizontal and vertical sizes of the CU before division, depending on the number of CUs created by division. For example, the horizontal and vertical sizes of the CU created by division may be half the horizontal size and half the vertical size of the CU before division.

A split CU can be recursively split into multiple CUs in the same manner. By recursive division, at least one of the horizontal and vertical sizes of the divided CU may be reduced compared to at least one of the horizontal and vertical sizes of the CU before division.

Division of the CU can be done recursively up to a predefined depth or predefined size.

For example, the depth of the CU may have a value of 0 to 3. The size of the CU can range from 64x64 to 8x8 depending on the depth of the CU.

For example, the depth of the LCU 310 may be 0, and the depth of the Smallest Coding Unit (SCU) may be a predefined maximum depth. Here, the LCU may be a CU with the maximum coding unit size as described above, and the SCU may be a CU with the minimum coding unit size.

Division may begin from the LCU 310, and the depth of the CU may increase by 1 whenever the horizontal and/or vertical size of the CU is reduced due to division.

For example, for each depth, an undivided CU may have a size of 2Nx2N. Additionally, in the case of a divided CU, a CU of 2Nx2N size may be divided into 4 CUs of NxN size. The size of N can be reduced by half each time the depth increases by 1.

Referring to FIG. 3, an LCU with a depth of 0 may be 64x64 pixels or a 64x64 block. 0 may be the minimum depth. A SCU with a depth of 3 may be 8x8 pixels or an 8x8 block. 3 may be the maximum depth. At this time, the CU of the 64x64 block, which is the LCU, can be expressed as depth 0. A CU in a 32x32 block can be expressed with a depth of 1. A CU in a 16x16 block can be expressed with a depth of 2. A CU of an 8x8 block, which is an SCU, can be expressed with a depth of 3.

Information about whether a CU is divided can be expressed through the division information of the CU. Segmentation information may be 1 bit of information. All CUs except SCU may include segmentation information. For example, the partition information value of a CU that is not divided may be a first value, and the partition information value of a divided CU may be a second value. When the division information indicates whether the CU is divided, the first value may be 0 and the second value may be 1.

For example, when one CU is divided into four CUs, the horizontal and vertical sizes of each of the four CUs created by division are half the horizontal size and half the vertical size of the CU before division, respectively. You can. When a CU of size 32x32 is divided into 4 CUs, the sizes of the 4 divided CUs may be 16x16. When one CU is divided into four CUs, it can be said that the CU is divided into a quad-tree form. In other words, it can be seen that quad-tree partitioning has been applied to the CU.

For example, when one CU is divided into two CUs, the horizontal or vertical size of each CU of the two CUs created by division is half the horizontal size or half the vertical size of the CU before division, respectively. You can. When a CU of size 32x32 is vertically divided into two CUs, the sizes of the two divided CUs may be 16x32. When a CU of size 32x32 is horizontally divided into two CUs, the sizes of the two divided CUs may be 32x16. When one CU is divided into two CUs, it can be said that the CU is divided in a binary-tree form. In other words, it can be seen that binary-tree partitioning has been applied to the CU.

For example, when one CU is divided into three CUs, three divided CUs can be created by dividing the horizontal or vertical size of the CU before division at a ratio of 1:2:1. For example, if a CU with a size of 16x32 is divided into three CUs in the horizontal direction, the three divided CUs may have sizes of 16x8, 16x16, and 16x8, respectively, from the top. For example, if a CU of size 32x32 is divided into three CUs in the vertical direction, the three divided CUs may have sizes of 8x32, 16x32, and 8x32, respectively, from the left. When one CU is divided into three CUs, it can be said that the CU is divided in a ternary-tree form. In other words, it can be seen that ternary-tree partitioning has been applied to the CU.

Both quad-tree type partitioning and binary-tree type partitioning were applied to the LCU 310 of FIG. 3.

In the encoding device 100, a Coding Tree Unit (CTU) of 64x64 size may be divided into a plurality of smaller CUs using a recursive Quad-Cree structure. One CU can be divided into four CUs with identical sizes. CUs can be divided recursively, and each CU can have a quad tree structure.

Through recursive partitioning of the CU, the optimal partitioning method that generates the minimum rate-distortion ratio can be selected.

The CTU 320 in FIG. 3 is an example of a CTU to which quad tree partitioning, binary tree partitioning, and ternary tree partitioning are all applied.

As described above, to partition a CTU, at least one of quad tree partitioning, binary tree partitioning, and ternary tree partitioning may be applied to the CTU. Partitions may be applied based on a specified priority.

For example, quad tree partitioning may be applied preferentially for CTU. A CU that can no longer be divided into a quad tree may correspond to a leaf node of the quad tree. The CU corresponding to the leaf node of the quad tree can be the root node of the binary tree and/or ternary tree. That is, the CU corresponding to the leaf node of the quad tree may be divided into a binary tree or a ternary tree, or may not be divided any further. At this time, quad tree division is not applied again to the CU created by applying binary tree division or ternary tree division to the CU corresponding to the leaf node of the quad tree, thereby preventing block division and/or signaling of block division information. It can be performed effectively.

The division of the CU corresponding to each node of the quad tree can be signaled using quad division information. Quad partition information with a first value (eg, “1”) may indicate that the CU is partitioned in a quad tree form. Quad partition information with a second value (eg, “0”) may indicate that the CU is not partitioned in a quad tree form. Quad split information may be a flag with a specified length (eg, 1 bit).

There may be no priority between binary tree partitioning and ternary tree partitioning. That is, the CU corresponding to the leaf node of the quad tree may be divided into a binary tree or a ternary tree. Additionally, the CU generated by binary tree partitioning or ternary tree partitioning may be partitioned again into a binary tree form or a ternary tree form, or may not be partitioned any further.

Partitioning when no priority exists between binary tree partitioning and ternary tree partitioning may be referred to as multi-type tree partitioning. In other words, the CU corresponding to the leaf node of the quad tree can become the root node of the multi-type tree. The division of the CU corresponding to each node of the multi-type tree may be signaled using at least one of information indicating whether the multi-type tree is divided, division direction information, and division tree information. To split the CU corresponding to each node of the multi-type tree, information indicating whether to split sequentially, split direction information, and split tree information may be signaled.

For example, information indicating whether a multi-type tree with a first value (eg, “1”) is split may indicate that the corresponding CU is split in the form of a multi-type tree. Information indicating whether a multi-type tree with a second value (eg, “0”) is divided may indicate that the corresponding CU is not divided into a multi-type tree.

When the CU corresponding to each node of the multi-type tree is split in the form of a multi-type tree, the corresponding CU may further include split direction information.

Splitting direction information may indicate the splitting direction of multi-type tree splitting. Division direction information with a first value (eg, “1”) may indicate that the corresponding CU is divided in the vertical direction. Division direction information with a second value (eg, “0”) may indicate that the corresponding CU is divided in the horizontal direction.

When the CU corresponding to each node of the multi-type tree is split into a multi-type tree, the corresponding CU may further include split tree information. Splitting tree information may indicate the tree used for multi-type tree splitting.

For example, split tree information with a first value (eg, “1”) may indicate that the corresponding CU is split in the form of a binary tree. Split tree information with a second value (eg, “0”) may indicate that the corresponding CU is split in a ternary tree form.

Here, each of the above-described information indicating whether to split, split tree information, and split direction information may be a flag with a specified length (eg, 1 bit).

At least one of the above-described quad split information, information indicating whether the multi-type tree is split, split direction information, and split tree information may be entropy encoded and/or entropy decoded. For entropy encoding/decoding of such information, information on a neighboring CU adjacent to the target CU can be used.

For example, the splitting form of the left CU and/or the upper CU (i.e., whether to split, splitting tree, and/or splitting direction) and the splitting form of the target CU may be considered highly likely to be similar to each other. Therefore, based on information on the neighboring CU, context information for entropy encoding and/or entropy decoding of information on the target CU may be derived. At this time, the information on the neighboring CU may include at least one of the neighboring CU's 1) quad split information, 2) information indicating whether the multi-type tree is split, 3) split direction information, and 4) split tree information.

As another embodiment, among binary tree partitioning and ternary tree partitioning, binary tree partitioning may be performed preferentially. That is, binary tree division is applied first, and the CU corresponding to the leaf node of the binary tree may be set as the root node of the ternary tree. In this case, quad tree division and binary tree division may not be performed on the CU corresponding to the node of the ternary tree.

A CU that is no longer split by quad tree splitting, binary tree splitting, and/or ternary tree splitting may become a unit of encoding, prediction, and/or transformation. That is, for prediction and/or transformation, the CU may no longer be split. Accordingly, a split structure and split information for splitting a CU into prediction units and/or transform units may not exist in the bitstream.

However, if the size of the CU that is the unit of division is larger than the size of the maximum conversion block, this CU may be recursively divided until the size of the CU becomes less than or equal to the size of the maximum conversion block. For example, if the size of the CU is 64x64 and the maximum conversion block size is 32x32, the CU may be divided into four 32x32 blocks for conversion. For example, if the size of the CU is 32x64 and the maximum conversion block size is 32x32, the CU may be divided into two 32x32 blocks for conversion.

In this case, information about whether the CU is divided for conversion may not be signaled separately. Without signaling, whether to split a CU may be determined by comparison between the horizontal size (and/or vertical size) of the CU and the horizontal size (and/or vertical size) of the maximum transform block. For example, if the horizontal size of the CU is larger than the horizontal size of the maximum transformation block, the CU may be divided vertically into two. Additionally, if the vertical size of the CU is larger than the vertical size of the maximum transformation block, the CU may be divided horizontally into two.

Information about the maximum size and/or minimum size of the CU and information about the maximum size and/or minimum size of the transform block may be signaled or determined at a higher level for the CU. For example, higher levels may be sequence level, picture level, tile level, tile group level, and slice level. For example, the minimum size of a CU may be determined to be 4x4. For example, the maximum size of a transform block may be determined to be 64x64. For example, the minimum size of the transform block may be determined to be 4x4.

Information about the minimum size of the CU corresponding to the leaf node of the quad tree (say, the quad tree minimum size) and/or the maximum depth of the path from the root node to the leaf node of the multi-type tree (say, the multi-type tree maximum Information about depth) may be signaled or determined at a higher level for the CU. For example, higher levels may be sequence level, picture level, slice level, tile group level, and tile level. Information about the quad tree minimum size and/or information about the multi-type tree maximum depth may be signaled or determined separately for each of the intra-slice and inter-slice.

Differential information about the size of the CTU and the maximum size of the transform block may be signaled or determined at a higher level for the CU. For example, higher levels may be sequence level, picture level, slice level, tile group level, and tile level. Information about the maximum size of the CU corresponding to each node of the binary tree (that is, the maximum size of the binary tree) may be determined based on the size and difference information of the CTU. The maximum size of the CU corresponding to each node of the ternary tree (that is, the maximum size of the ternary tree) may have different values depending on the type of slice. For example, within an intra slice, the maximum ternary tree size may be 32x32. Additionally, for example, within an inter slice, the maximum ternary tree size may be 128x128. For example, the minimum size of a CU corresponding to each node in a binary tree (say, the binary tree minimum size) and/or the minimum size of a CU corresponding to each node in a ternary tree (say, the ternary tree minimum size) is Can be set to the minimum size.

As another example, the binary tree maximum size and/or the ternary tree maximum size may be signaled or determined at the slice level. Additionally, the binary tree minimum size and/or ternary tree minimum size may be signaled or determined at the slice level.

Based on the various block sizes and various depths described above, quad split information, information indicating whether the multi-type tree is split, split tree information, and/or split direction information may or may not exist in the bitstream.

For example, if the size of the CU is not larger than the quad tree minimum size, the CU may not include quad partition information, and the quad partition information for the CU may be inferred as the second value.

For example, the size (horizontal and vertical size) of the CU corresponding to a node in a multi-type tree is larger than the binary tree maximum size (horizontal size and vertical size) and/or the ternary tree maximum size (horizontal size and vertical size). For larger cases, the CU may not be partitioned into binary and/or ternary tree form. According to this decision method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, the size (horizontal and vertical size) of the CU corresponding to the node of the multi-type tree is equal to the minimum size (horizontal and vertical size) of the binary tree, or the size of the CU (horizontal and vertical size) is equal to the minimum size (horizontal and vertical size) of the binary tree. If equal to twice the minimum size (horizontal and vertical sizes), the CU may not be partitioned into binary tree form and/or ternary tree form. According to this decision method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value. This is because, when dividing a CU into a binary tree form and/or a ternary tree form, a CU smaller than the minimum binary tree size and/or the minimum ternary tree size is generated.

Alternatively, binary tree partitioning or ternary tree partitioning may be limited based on the size of the virtual pipeline data unit (i.e., pipeline buffer size). For example, if a CU is split into sub-CUs that do not fit the pipeline buffer size by binary tree partitioning or ternary tree partitioning, binary tree partitioning or ternary tree partitioning may be limited. The pipeline buffer size may be equal to the size of the maximum conversion block (e.g., 64X64).

For example, when the pipeline buffer size is 64X64, the following partitions may be limited.

- ternary tree split for NxM (N and/or M is 128) CUs

- Horizontally directed binary tree split for 128xN (N <= 64) CUs

- Vertically oriented binary tree partitioning for Nx128 (N <= 64) CUs

Alternatively, if the depth within the multi-type tree of the CU corresponding to the node of the multi-type tree is equal to the maximum depth of the multi-type tree, the CU may not be divided into a binary tree form and/or a ternary tree form. According to this decision method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, a multi-type tree only if at least one of vertical binary tree partitioning, horizontal binary tree partitioning, vertical ternary tree partitioning, and horizontal ternary tree partitioning is possible for the CU corresponding to the node of the multi-type tree. Information indicating whether to divide may be signaled. Otherwise, the CU may not be partitioned into binary tree form and/or ternary tree form. According to this decision method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, split direction information only if both vertical binary tree splitting and horizontal binary tree splitting are possible for the CU corresponding to the node of the multi-type tree, or both vertical ternary tree splitting and horizontal ternary tree splitting are possible. can be signaled. Otherwise, the division direction information may not be signaled and may be inferred as a value indicating the direction in which the CU can be divided.

Alternatively, split tree information only if both vertical binary tree splitting and vertical ternary tree splitting are possible for the CU corresponding to the node of the multi-type tree, or both horizontal binary tree splitting and horizontal ternary tree splitting are possible. can be signaled. Otherwise, the split tree information may not be signaled and may be inferred as a value indicating a tree applicable to splitting the CU.

Figure 4 is a diagram showing the form of a prediction unit that a coding unit can include.

Among the CUs divided from the LCU, CUs that are no longer divided may be divided into one or more prediction units (PUs).

PU may be the basic unit for prediction. PU can be encoded and decoded in any one of skip mode, inter mode, and intra mode. PU can be divided into various forms depending on each mode. For example, the target block described above with reference to FIG. 1 and the target block described with reference to FIG. 2 may be a PU.

A CU may not be divided into PUs. If the CU is not divided into PUs, the size of the CU and the size of the PU may be the same.

In skip mode, there may be no partitions within the CU. In skip mode, 2Nx2N mode 410 in which the sizes of PU and CU are the same without division can be supported.

In inter mode, eight partition types can be supported within the CU. For example, in inter mode, 2Nx2N mode (410), 2NxN mode (415), Nx2N mode (420), NxN mode (425), 2NxnU mode (430), 2NxnD mode (435), nLx2N mode (440), and nRx2N Mode 445 may be supported.

In intra mode, 2Nx2N mode 410 and NxN mode 425 may be supported.

In 2Nx2N mode 410, a PU with a size of 2Nx2N can be encoded. A PU of size 2Nx2N may mean a PU of the same size as the size of the CU. For example, a PU of size 2Nx2N may have sizes of 64x64, 32x32, 16x16 or 8x8.

In NxN mode 425, PUs of NxN size can be encoded.

For example, in intra prediction, when the size of a PU is 8x8, four divided PUs can be encoded. The size of the divided PU may be 4x4.

When the PU is encoded by intra mode, the PU may be encoded using one intra prediction mode among a plurality of intra prediction modes. For example, High Efficiency Video Coding (HEVC) technology can provide 35 intra prediction modes, and a PU can be encoded with one intra prediction mode among the 35 intra prediction modes.

Which of the 2Nx2N mode 410 and NxN mode 425 will be used to encode the PU can be determined by the rate-distortion cost.

The encoding device 100 can perform an encoding operation on a PU of size 2Nx2N. Here, the encoding operation may be encoding the PU in each of a plurality of intra prediction modes that the encoding device 100 can use. Through encoding operations, the optimal intra prediction mode for a PU of size 2Nx2N can be derived. The optimal intra prediction mode may be an intra prediction mode that generates the minimum rate-distortion cost for encoding a PU of 2Nx2N size among a plurality of intra prediction modes that the encoding device 100 can use.

Additionally, the encoding device 100 may sequentially perform an encoding operation on each PU of the NxN divided PUs. Here, the encoding operation may be encoding the PU in each of a plurality of intra prediction modes that the encoding device 100 can use. The optimal intra prediction mode for a PU of NxN size can be derived through encoding operations. The optimal intra prediction mode may be an intra prediction mode that generates the minimum rate-distortion cost for encoding an NxN sized PU among a plurality of intra prediction modes that the encoding device 100 can use.

The encoding device 100 may determine which of the 2Nx2N sized PUs and NxN sized PUs to encode based on comparison of the rate-distortion costs of the 2Nx2N sized PU and the rate-distortion costs of the NxN sized PUs.

One CU can be divided into one or more PUs, and a PU can also be divided into multiple PUs.

For example, when one PU is divided into four PUs, the horizontal and vertical sizes of each of the four PUs created by division are half the horizontal size and half the vertical size of the PU before division, respectively. You can. When a PU of size 32x32 is divided into 4 PUs, the sizes of the 4 divided PUs may be 16x16. When one PU is divided into four PUs, it can be said that the PU is divided into a quad-tree form.

For example, when one PU is divided into two PUs, the horizontal or vertical size of each PU of the two PUs created by division is half the horizontal size or half the vertical size of the PU before division, respectively. You can. When a PU of size 32x32 is vertically divided into two PUs, the sizes of the two divided PUs may be 16x32. When a PU of size 32x32 is horizontally divided into two PUs, the sizes of the two divided PUs may be 32x16. When one PU is divided into two PUs, it can be said that the PU is divided into a binary-tree form.

Figure 5 is a diagram showing the form of a conversion unit that can be included in a coding unit.

Transform Unit (TU) may be a basic unit used for the processes of transformation, quantization, inverse transformation, inverse quantization, entropy encoding, and entropy decoding within the CU.

TU may have a square or rectangular shape. The shape of the TU may be determined depending on the size and/or shape of the CU.

Among the CUs divided from the LCU, CUs that are no longer divided into CUs may be divided into one or more TUs. At this time, the division structure of the TU may be a quad-tree structure. For example, as shown in FIG. 5, one CU 510 may be divided one or more times according to a quad-tree structure. Through division, one CU 510 can be composed of TUs of various sizes.

If one CU is divided more than two times, the CU can be viewed as being divided recursively. Through partitioning, one CU can be composed of TUs with various sizes.

Alternatively, one CU may be divided into one or more TUs based on the number of vertical lines and/or horizontal lines dividing the CU.

A CU may be divided into symmetric TUs or may be divided into asymmetric TUs. For division into asymmetric TUs, information about the size and/or shape of the TU may be signaled from the encoding device 100 to the decoding device 200. Alternatively, the size and/or shape of the TU may be derived from information about the size and/or shape of the CU.

A CU may not be divided into TUs. If the CU is not divided into TUs, the size of the CU and the size of the TU may be the same.

One CU may be divided into one or more TUs, and a TU may also be divided into multiple TUs.

For example, when one TU is split into four TUs, the horizontal and vertical sizes of each of the four TUs created by the split are half the horizontal size and half the vertical size of the TU before splitting, respectively. You can. When a TU of size 32x32 is divided into 4 TUs, the sizes of the 4 divided TUs may be 16x16. When one TU is divided into four TUs, it can be said that the TU is divided into a quad-tree form.

For example, when one TU is split into two TUs, the horizontal or vertical size of each TU of the two TUs created by splitting is half the horizontal size or half the vertical size of the TU before splitting, respectively. You can. When a TU of size 32x32 is vertically divided into two TUs, the sizes of the two divided TUs may be 16x32. When a TU of size 32x32 is horizontally divided into two TUs, the sizes of the two divided TUs may be 32x16. When one TU is divided into two TUs, it can be said that the TU is divided into a binary-tree form.

The CU may be divided in a manner other than that shown in FIG. 5.

For example, one CU can be divided into three CUs. The horizontal or vertical size of the three divided CUs may be 1/4, 1/2, and 1/4 of the horizontal or vertical size of the CU before division, respectively.

For example, when a 32x32 CU is vertically divided into 3 CUs, the sizes of the 3 divided CUs may be 8x32, 16x32, and 8x32, respectively. In this way, when one CU is divided into three CUs, the CU can be viewed as being divided in the form of a ternary tree.

One of the exemplified quad tree-type partitioning, binary tree-type partitioning, and ternary tree-type partitioning may be applied for partitioning the CU, and a plurality of partitioning methods may be combined together and used for partitioning the CU. . At this time, the case where a plurality of partitioning methods are used in combination can be referred to as partitioning in the form of a composite tree.

Figure 6 shows division of a block according to an example.

In the process of encoding and/or decoding an image, the target block may be divided as shown in FIG. 6. For example, the target block may be a CU.

To split the target block, an indicator indicating splitting information may be signaled from the encoding device 100 to the decoding device 200. Splitting information may be information indicating how the target block is divided.

Splitting information includes split flag (hereinafter referred to as “split_flag”), quad-binary flag (hereinafter referred to as “QB_flag”), quad tree flag (hereinafter referred to as “quadtree_flag”), and binary tree flag (hereinafter referred to as “binarytree_flag”). It may be one or more of a binary type flag (hereinafter denoted as "Btype_flag").

split_flag may be a flag indicating whether the block is split. For example, a value of 1 in split_flag may indicate that the block is split. A value of 0 for split_flag may indicate that the block is not split.

QB_flag may be a flag indicating whether the block is divided into a quad tree format or a binary tree format. For example, a value of 0 for QB_flag may indicate that the block is divided into a quad tree format. A value of 1 for QB_flag may indicate that the block is divided into a binary tree form. Alternatively, the value of QB_flag 0 may indicate that the block is divided into a binary tree form. A value of 1 in QB_flag may indicate that the block is divided into a quad tree format.

quadtree_flag may be a flag indicating whether the block is divided into a quad tree format. For example, a value of 1 in quadtree_flag may indicate that the block is divided into a quad tree format. A value of 0 for quadtree_flag may indicate that the block is not divided into a quad tree format.

binarytree_flag may be a flag indicating whether the block is divided in binary tree form. For example, a value of 1 in binarytree_flag may indicate that the block is split into a binary tree. A value of 0 for binarytree_flag may indicate that the block is not divided into a binary tree form.

Btype_flag may be a flag indicating whether the block is divided into vertical division or horizontal division when the block is divided into binary tree form. For example, a value of 0 in Btype_flag may indicate that the block is divided in the horizontal direction. A value of 1 in Btype_flag may indicate that the block is divided in the vertical direction. Alternatively, the value 0 of Btype_flag may indicate that the block is divided in the vertical direction. A value of 1 in Btype_flag may indicate that the block is divided in the horizontal direction.

For example, partition information for the block of FIG. 6 can be derived by signaling at least one of quadtree_flag, binarytree_flag, and Btype_flag as shown in Table 1 below.

[Table 1]

For example, split information for the block of FIG. 6 can be derived by signaling at least one of split_flag, QB_flag, and Btype_flag as shown in Table 2 below.

[Table 2]

The partitioning method may be limited to quad trees only, or only binary trees, depending on the size and/or shape of the block. When this restriction is applied, split_flag may be a flag indicating whether to split into a quad tree form or a flag indicating whether to split into a binary tree form. The size and shape of the block can be derived according to the depth information of the block, and the depth information can be signaled from the encoding device 100 to the decoding device 200.

If the size of the block falls within a specified range, only quad tree-type division may be possible. For example, the specified range may be defined by at least one of the maximum block size and minimum block size for which only quad tree-type division is possible.

Information indicating the maximum block size and/or minimum block size for which only quad tree-type division is possible may be signaled from the encoding device 100 to the decoding device 200 through a bitstream. Additionally, this information may be signaled for at least one unit among video, sequence, picture, parameter, tile group, and slice (or segment).

Alternatively, the maximum block size and/or minimum block size may be fixed sizes predefined in the encoding device 100 and the decoding device 200. For example, if the block size is 64x64 or larger and 256x256 or smaller, only quad tree-type division may be possible. In this case, split_flag may be a flag indicating whether to split into quad tree form.

If the block size is larger than the maximum conversion block size, only quad tree-type division may be possible. At this time, the divided block may be at least one of CU and TU.

In this case, split_flag may be a flag indicating whether to split into quad tree form.

If the block size falls within a specified range, only binary tree or ternary tree division may be possible. Here, for example, the specified range may be defined by at least one of the maximum block size and minimum block size for which only division in the form of a binary tree or a ternary tree is possible.

Information indicating the maximum block size and/or minimum block size for which only binary tree-type splitting or ternary tree-type splitting is possible may be signaled from the encoding device 100 to the decoding device 200 through a bitstream. Additionally, this information may be signaled for at least one unit among sequence, picture, and slice (or segment).

Alternatively, the maximum block size and/or minimum block size may be fixed sizes predefined in the encoding device 100 and the decoding device 200. For example, if the block size is 8x8 or larger and 16x16 or smaller, only division in the form of a binary tree may be possible. In this case, split_flag may be a flag indicating whether to split in binary tree form or ternary tree form.

The above-described description of quad tree-type partitioning can be equally applied to binary tree-type and/or ternary-tree form partitioning.

Splitting of a block may be limited by previous splitting. For example, when a block is divided into a specified binary tree form and a plurality of divided blocks are created, each divided block can be further divided only into the specified tree form. Here, the specified tree form may be at least one of a binary tree form, a ternary tree form, and a quad tree form.

If the horizontal or vertical size of the divided block corresponds to a size that cannot be further divided, the above-described indicator may not be signaled.

Figure 7 is a diagram for explaining an embodiment of the intra prediction process.

Arrows from the center to the outside of the graph of FIG. 7 may indicate prediction directions of directional intra prediction modes. Additionally, numbers displayed close to the arrows may represent an example of a mode value assigned to the intra prediction mode or the prediction direction of the intra prediction mode.

In FIG. 7, the number 0 may represent Planar mode, which is a non-directional intra prediction mode. The number 1 may represent DC mode, which is a non-directional intra prediction mode.

Intra encoding and/or decoding may be performed using reference samples of neighboring blocks of the target block. The neighboring block may be a reconstructed neighboring block. A reference sample may refer to a neighboring sample.

For example, intra encoding and/or decoding may be performed using the value or coding parameter of a reference sample included in the reconstructed neighboring block.

The encoding device 100 and/or the decoding device 200 may generate a prediction block by performing intra prediction on the target block based on information on samples in the target image. When performing intra prediction, the encoding device 100 and/or the decoding device 200 may generate a prediction block for the target block by performing intra prediction based on information on samples in the target image. When performing intra prediction, the encoding device 100 and/or the decoding device 200 may perform directional prediction and/or non-directional prediction based on at least one reconstructed reference sample.

A prediction block may refer to a block generated as a result of performing intra prediction. A prediction block may correspond to at least one of CU, PU, and TU.

The unit of the prediction block may be the size of at least one of CU, PU, and TU. The prediction block may have a square shape with a size of 2Nx2N or NxN. NxN sizes can include 4x4, 8x8, 16x16, 32x32, and 64x64.

Alternatively, the prediction block may be a square-shaped block with a size of 2x2, 4x4, 8x8, 16x16, 32x32, or 64x64, or a rectangular block with a size of 2x8, 4x8, 2x16, 4x16, and 8x16. there is.

Intra prediction may be performed according to the intra prediction mode for the target block. The number of intra prediction modes that a target block can have may be a predefined fixed value or a value determined differently depending on the properties of the prediction block. For example, properties of the prediction block may include the size of the prediction block and the type of the prediction block. Additionally, properties of a prediction block may indicate coding parameters for the prediction block.

For example, the number of intra prediction modes may be fixed to N regardless of the size of the prediction block. Or, for example, the number of intra prediction modes may be 3, 5, 9, 17, 34, 35, 36, 65, 67, or 95.

The intra prediction mode may be a non-directional mode or a directional mode.

For example, an intra prediction mode may include 2 undirectional modes and 65 directional modes, corresponding to numbers 0 to 66 shown in FIG. 7 .

For example, when a specified intra prediction method is used, the intra prediction mode may include 2 undirectional modes and 93 directional modes, corresponding to numbers -14 to 80 shown in FIG. 7.

The two non-directional modes may include DC mode and Planar mode.

The directional mode may be a prediction mode with a specific direction or a specific angle. Directional mode may also be referred to as an argular mode.

The intra prediction mode may be expressed by at least one of a mode number, mode value, mode angle, and mode direction. That is to say, the terms “(mode) number of intra prediction mode”, “(mode) value of intra prediction mode”, “(mode) angle of intra prediction mode” and “(mode) direction of intra prediction mode” have the same meaning. can be used, and can be used interchangeably.

The number of intra prediction modes may be M. M may be 1 or more. In other words, the number of intra prediction modes may be M, including the number of non-directional modes and the number of directional modes.

The number of intra prediction modes may be fixed to M regardless of the size and/or color component of the block. For example, the number of intra prediction modes may be fixed to either 35 or 67, regardless of the block size.

Alternatively, the number of intra prediction modes may vary depending on the shape, size, and/or type of color component of the block.

For example, in FIG. 7, directional prediction modes shown in dotted lines can only be applied to prediction for non-square blocks.

For example, as the block size increases, the number of intra prediction modes may increase. Alternatively, as the block size increases, the number of intra prediction modes may decrease. If the block size is 4x4 or 8x8, the number of intra prediction modes may be 67. If the block size is 16x16, the number of intra prediction modes may be 35. If the block size is 32x32, the number of intra prediction modes may be 19. If the block size is 64x64, the number of intra prediction modes may be 7.

For example, the number of intra prediction modes may vary depending on whether the color component is a luma signal or a chroma signal. Alternatively, the number of intra prediction modes of the luma component block may be greater than the number of intra prediction modes of the chroma component block.

For example, in the case of vertical mode with a mode value of 50, prediction may be performed in the vertical direction based on the pixel value of the reference sample. For example, in the case of horizontal mode where the mode value is 18, prediction may be performed in the horizontal direction based on the pixel value of the reference sample.

Even in the case of a directional mode other than the above-described mode, the encoding device 100 and the decoding device 200 can perform intra prediction on the target unit using a reference sample according to the angle corresponding to the directional mode.

The intra prediction mode located to the right of the vertical mode may be named vertical-right mode. The intra prediction mode located below the horizontal mode may be named the horizontal-below mode. For example, in Figure 7, intra prediction modes with mode values one of 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, and 66 are vertical These may be the right modes. Intra prediction modes with mode values of one of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, and 17 may be horizontal bottom modes.

Non-directional modes may include DC mode and planar mode. For example, the mode value of DC mode may be 1. The mode value of the planner mode may be 0.

Directional modes may include angular modes. Among the plurality of intra prediction modes, the remaining modes except DC mode and planner mode may be directional modes.

When the intra prediction mode is DC mode, a prediction block may be generated based on the average of pixel values of a plurality of reference samples. For example, the pixel value of the prediction block may be determined based on the average of pixel values of a plurality of reference samples.

The number of intra prediction modes described above and the mode value of each intra prediction mode may be merely exemplary. The number of intra prediction modes described above and the mode value of each intra prediction mode may be defined differently depending on embodiment, implementation, and/or need.

In order to perform intra prediction on the target block, a step may be performed to check whether samples included in the reconstructed neighboring block can be used as reference samples of the target block. If there is a sample among the samples in the neighboring block that cannot be used as a reference sample for the target block, a value generated by copying and/or interpolation using at least one sample value among the samples included in the reconstructed neighboring block. This can be replaced with the sample value of a sample that cannot be used as a reference sample. If the value generated by copying and/or interpolation is replaced with the sample value of the sample, the sample can be used as a reference sample of the target block.

When intra prediction is used, a filter may be applied to at least one of a reference sample or a prediction sample based on at least one of the intra prediction mode and the size of the target block.

The type of filter applied to at least one of the reference sample and the prediction sample may vary depending on at least one of the intra prediction mode of the target block, the size of the target block, and the shape of the target block. The type of filter can be classified according to one or more of the length of the filter tap, the value of the filter coefficient, and the filter strength. The length of the above filter tabs may mean the number of filter tabs. Additionally, the number of filter tabs may mean the length of the filter.

When the intra prediction mode is planar mode, when generating the prediction block of the target block, depending on the location of the prediction target sample in the prediction block, the upper reference sample of the target sample, the left reference sample of the target sample, and the upper right reference sample of the target block And the sample value of the prediction target sample may be generated using the weighted sum (weight-sum) of the lower left reference sample of the target block.

When the intra prediction mode is DC mode, when generating the prediction block of the target block, the average value of the top reference samples and the left reference samples of the target block can be used. Additionally, filtering using values of reference samples may be performed on specified rows or specified columns within the target block. The rows specified may be one or more top rows adjacent to the reference sample. The specified columns may be one or more left columns adjacent to the reference sample.

When the intra prediction mode is a directional mode, a prediction block may be generated using the top reference sample, left reference sample, top right reference sample, and/or bottom left reference sample of the target block.

Real-valued interpolation may be performed to generate the prediction samples described above.

The intra prediction mode of the target block may be predicted from the intra prediction mode of the target block's neighboring block, and information used for prediction may be entropy encoded/decoded.

For example, if the intra prediction modes of the target block and the neighboring block are the same, it may be signaled that the intra prediction modes of the target block and the neighboring block are the same using a predefined flag.

For example, an indicator indicating an intra prediction mode that is the same as the intra prediction mode of the target block among the intra prediction modes of a plurality of neighboring blocks may be signaled.

If the intra prediction modes of the target block and the neighboring block are different from each other, information on the intra prediction mode of the target block may be encoded and/or decoded using entropy coding and/or decoding.

Figure 8 is a diagram for explaining reference samples used in the intra prediction process.

The reconstructed reference samples used for intra prediction of the target block are below-left reference samples, left reference samples, above-left corner reference samples, and above reference samples. and above-right reference samples, etc.

For example, left reference samples may refer to reconstructed reference pixels adjacent to the left side of the target block. Top reference samples may refer to reconstructed reference pixels adjacent to the top of the target block. The upper left corner reference sample may refer to a reconstructed reference pixel located at the upper left corner of the target block. Additionally, the lower left reference samples may refer to a reference sample located at the bottom of the left sample line among samples located on the same line as the left sample line composed of left reference samples. The upper right reference samples may refer to reference samples located to the right of the upper pixel line among samples located on the same line as the upper sample line composed of upper reference samples.

When the size of the target block is NxN, the number of bottom left reference samples, left reference samples, top reference samples, and top right reference samples may each be N.

A prediction block may be generated through intra prediction for the target block. Generating a prediction block may include determining values of pixels of the prediction block. The sizes of the target block and prediction block may be the same.

The reference sample used for intra prediction of the target block may vary depending on the intra prediction mode of the target block. The direction of the intra prediction mode may indicate a dependency relationship between reference samples and pixels of the prediction block. For example, the value of a specified reference sample can be used as the value of one or more specified pixels of the prediction block. In this case, the specified reference sample and one or more specified pixels of the prediction block may be samples and pixels designated by a straight line in the direction of the intra prediction mode. In other words, the value of the specified reference sample can be copied to the value of the pixel located in the reverse direction of the intra prediction mode. Alternatively, the pixel value of the prediction block may be the value of a reference sample located in the direction of the intra prediction mode based on the position of the pixel.

For example, when the intra prediction mode of the target block is vertical mode, top reference samples can be used for intra prediction. When the intra prediction mode is a vertical mode, the pixel value of the prediction block may be the value of a reference sample located vertically above the position of the pixel. Therefore, top reference samples adjacent to the top of the target block can be used for intra prediction. Additionally, the values of pixels in one row of the prediction block may be the same as the values of the upper reference samples.

For example, when the intra prediction mode of the target block is horizontal mode, left reference samples can be used for intra prediction. When the intra prediction mode is a horizontal mode, the pixel value of the prediction block may be the value of a reference sample located horizontally to the left of the pixel. Therefore, left reference samples adjacent to the left of the target block can be used for intra prediction. Additionally, the values of pixels in one column of the prediction block may be the same as the values of the left reference samples.

For example, when the mode value of the intra prediction mode of the target block is 34, at least some of the left reference samples, the top left corner reference sample, and at least some of the top reference samples may be used for intra prediction. When the mode value of the intra prediction mode is 34, the pixel value of the prediction block may be the value of a reference sample located diagonally to the upper left with respect to the pixel.

Additionally, when an intra prediction mode with a mode value of one of 52 to 66 is used, at least some of the upper right reference samples may be used for intra prediction.

Additionally, when an intra prediction mode with a mode value of one of 2 to 17 is used, at least some of the lower left reference samples may be used for intra prediction.

Additionally, when an intra prediction mode with a mode value of one of 19 to 49 is used, the upper left corner reference sample can be used for intra prediction.

The reference sample used to determine the pixel value of one pixel of the prediction block may be one or two or more.

As described above, the pixel value of the pixel of the prediction block may be determined according to the location of the pixel and the location of the reference sample indicated by the direction of the intra prediction mode. If the position of the reference sample indicated by the pixel position and the direction of the intra prediction mode is an integer position, the value of one reference sample indicated by the integer position may be used to determine the pixel value of the pixel of the prediction block.

If the position of the reference sample indicated by the pixel position and the direction of the intra prediction mode is not an integer position, an interpolated reference sample can be generated based on the two reference samples closest to the position of the reference sample. there is. The value of the interpolated reference sample can be used to determine the pixel value of the pixel of the prediction block. In other words, when the position of the reference sample indicated by the position of the pixel of the prediction block and the direction of the intra prediction mode indicates the gap between two reference samples, an interpolated value is generated based on the values of the two samples. You can.

The prediction block generated by prediction may not be identical to the original target block. In other words, there may be a prediction error, which is a difference between the target block and the prediction block, and there may also be a prediction error between the pixels of the target block and the pixels of the prediction block.

Hereinafter, the terms “difference”, “error” and “residual” may be used with the same meaning and may be used interchangeably.

For example, in the case of directional intra prediction, the larger the distance between the pixel of the prediction block and the reference sample, the larger the prediction error may occur. Discontinuity may occur between the prediction block and neighboring blocks generated due to such prediction errors, etc.

Filtering on prediction blocks may be used to reduce prediction error. Filtering may be adaptively applying a filter to an area considered to have a large prediction error among prediction blocks. For example, an area considered to have a large prediction error may be the boundary of a prediction block. Additionally, depending on the intra-prediction mode, the area considered to have a large prediction error among prediction blocks may be different, and the characteristics of the filter may be different.

As shown in FIG. 8, at least one of reference lines 0 to 3 may be used for intra prediction of the target block.

Each reference line in FIG. 8 may represent a reference sample line including one or more reference samples. The smaller the reference line number, the closer the reference sample line may be to the target block.

Instead of being obtained from a reconstructed neighboring block, the samples of segment A and segment F may be obtained through padding using the closest samples of segment B and segment E, respectively.

Index information indicating a reference sample line to be used for intra prediction of the target block may be signaled. Index information may indicate a reference sample line used for intra prediction of a target block among a plurality of reference sample lines. For example, index information may have a value between 0 and 3.

If the upper boundary of the target block is the boundary of the CTU, only reference sample line 0 may be available. Therefore, in this case, index information may not be signaled. If a reference sample line other than reference sample line 0 is used, filtering on the prediction block, which will be described later, may not be performed.

In the case of inter-color intra prediction, a prediction block for the target block of the second color component may be generated based on the corresponding reconstructed block of the first color component.

For example, the first color component may be a luma component, and the second color component may be a chroma component.

For intra prediction between color components, parameters of a linear model between the first color component and the second color component may be derived based on the template.

The template may include a top reference sample and/or a left reference sample of the target block, and may include a top reference sample and/or a left reference sample of the reconstructed block of the first color component corresponding to these reference samples. there is.

For example, the parameters of a linear model are 1) the value of the sample of the first color component that has the maximum value among the samples in the template, 2) the value of the sample of the second color component corresponding to this sample of the first color component, 3) the value of the sample of the first color component having the minimum value among the samples in the template, and 4) the value of the sample of the second color component corresponding to the sample of the first color component.

Once the parameters of the linear model are derived, a prediction block for the target block can be generated by applying the corresponding reconstructed block to the linear model.

Depending on the video format, subsampling may be performed on neighboring samples of the reconstructed block of the first color component and the corresponding reconstructed block. For example, if 1 sample of the second color component corresponds to 4 samples of the first color component, 1 corresponding sample can be calculated by subsampling the 4 samples of the first color component. there is. When subsampling is performed, derivation of parameters of a linear model and intra prediction between color components can be performed based on the subsampled corresponding samples.

Whether to perform intra prediction between color components and/or the range of the template may be signaled as an intra prediction mode.

The target block may be divided into 2 or 4 sub-blocks in the horizontal and/or vertical directions.

The divided sub-blocks can be sequentially reconstructed. That is, as intra prediction is performed on the sub-block, a sub-prediction block for the sub-block may be generated. Additionally, as inverse quantization and/or inverse transformation is performed on the sub-block, a sub-residual block for the sub-block may be generated. A reconstructed sub-block can be generated by adding the sub-prediction block to the sub-residual block. The reconstructed subblock can be used as a reference sample for intra prediction of a later-ranked subblock.

A subblock may be a block containing a specified number (eg, 16) or more samples. Therefore, for example, if the target block is an 8x4 block or a 4x8 block, the target block may be divided into two sub-blocks. Additionally, if the target block is a 4x4 block, the target block cannot be divided into sub-blocks. If the target block has a different size, the target block may be divided into 4 sub-blocks.

Information regarding whether intra prediction based on these subblocks is performed and/or the division direction (horizontal or vertical direction) may be signaled.

Such sub-block-based intra prediction may be limited to being performed only when using reference sample line 0. When sub-block-based intra prediction is performed, filtering on the prediction block, which will be described later, may not be performed.

A final prediction block can be generated by performing filtering on the prediction block generated by intra prediction.

Filtering may be performed by applying a specific weight to the filtering target sample, left reference sample, top reference sample, and/or top left reference sample that are the objects of filtering.

Weights and/or reference samples (or ranges of reference samples or positions of reference samples, etc.) used for filtering may be determined based on at least one of block size, intra prediction mode, and location within the prediction block of the sample to be filtered. there is.

For example, filtering may be performed only for specified intra prediction modes (eg, DC mode, planar mode, vertical mode, horizontal mode, diagonal mode, and/or adjacent diagonal mode).

The adjacent diagonal mode may be a mode with a number in which k is added to the number of the diagonal mode, or it may be a mode with a number in which k is subtracted from the number of the diagonal mode. That is, the number of adjacent diagonal modes may be the sum of the number of diagonal modes and k, or the difference between the number of diagonal modes and k. For example, k may be a positive integer of 8 or less.

The intra prediction mode of the target block may be derived using the intra prediction mode of a neighboring block existing around the target block, and this derived intra prediction mode may be entropy encoded and/or entropy decoded.

For example, if the intra prediction mode of the target block and the intra prediction mode of the neighboring block are the same, information that the intra prediction mode of the target block and the intra prediction mode of the neighboring block are the same may be signaled using the specified flag information. .

Additionally, for example, indicator information about a neighboring block having the same intra prediction mode as the intra prediction mode of the target block among the intra prediction modes of a plurality of neighboring blocks may be signaled.

For example, if the intra prediction mode of the target block and the intra prediction mode of the neighboring block are different from each other, entropy coding and/or entropy decoding based on the intra prediction mode of the neighboring block are performed to obtain information about the intra prediction mode of the target block. Entropy encoding and/or entropy decoding may be performed.

Figure 9 is a diagram for explaining an embodiment of the inter prediction process.

The square shown in FIG. 9 may represent an image (or picture). Additionally, in FIG. 9, the arrow may indicate the prediction direction. An arrow from the first picture to the second picture may indicate that the second picture refers to the first picture. That is, the image can be encoded and/or decoded according to the prediction direction.

Each image can be classified into I picture (Intra Picture), P picture (Uni-prediction Picture), and B picture (Bi-prediction Picture) depending on the encoding type. Each picture may be encoded and/or decoded according to the encoding type of each picture.

If the target image that is the target of encoding is an I picture, the target image can be encoded using data within the image itself without inter prediction referring to other images. For example, an I picture can be encoded only with intra prediction.

When the target image is a P picture, the target image can be encoded through inter prediction using only reference pictures that exist in one direction. Here, unidirectional can be forward or reverse.

When the target image is a B picture, the target image may be encoded through inter prediction using reference pictures existing in both directions or inter prediction using reference pictures existing in one of the forward and reverse directions. Here, the two directions can be forward and reverse.

P pictures and B pictures that are encoded and/or decoded using a reference picture may be considered images for which inter prediction is used.

Below, inter prediction in inter mode according to the embodiment is described in detail.

Inter prediction or motion compensation can be performed using reference images and motion information.

In inter mode, the encoding device 100 may perform inter prediction and/or motion compensation for the target block. The decoding device 200 may perform inter prediction and/or motion compensation corresponding to the inter prediction and/or motion compensation in the encoding device 100 on the target block.

Motion information about the target block may be derived by each of the encoding device 100 and the decoding device 200 during inter prediction. The motion information may be derived using motion information of a reconstructed neighboring block, motion information of a call block, and/or motion information of a block adjacent to the call block.

For example, the encoding device 100 or the decoding device 200 performs prediction and/or motion compensation by using motion information of a spatial candidate and/or temporal candidate as motion information of the target block. It can be done. The target block may refer to PU and/or PU partition.

The spatial candidate may be a reconstructed block that is spatially adjacent to the target block.

The temporal candidate may be a reconstructed block corresponding to a target block in an already reconstructed collocated picture (col picture).

In inter prediction, the encoding device 100 and the decoding device 200 can improve encoding efficiency and decoding efficiency by using motion information of spatial candidates and/or temporal candidates. The motion information of the spatial candidate may be referred to as spatial motion information. The motion information of the temporal candidate may be referred to as temporal motion information.

Hereinafter, the motion information of the spatial candidate may be the motion information of the PU including the spatial candidate. The motion information of the temporal candidate may be motion information of a PU including the temporal candidate. The motion information of the candidate block may be motion information of the PU including the candidate block.

Inter prediction can be performed using a reference picture.

A reference picture may be at least one of a picture before the target picture or a picture after the target picture. A reference picture may refer to an image used for prediction of a target block.

In inter prediction, an area within a reference picture can be specified by using a reference picture index (or refIdx) indicating the reference picture and a motion vector to be described later. Here, a specified area within the reference picture may represent a reference block.

Inter prediction can select a reference picture and select a reference block corresponding to the target block within the reference picture. Additionally, inter prediction can generate a prediction block for the target block using the selected reference block.

Motion information may be derived during inter prediction by each of the encoding device 100 and the decoding device 200.

The spatial candidate may be a block that 1) exists in the target picture, 2) has already been reconstructed through encoding and/or decoding, and 3) is adjacent to the target block or located at a corner of the target block. Here, the block located at the corner of the target block may be a block vertically adjacent to a neighboring block horizontally adjacent to the target block, or a block horizontally adjacent to a neighboring block vertically adjacent to the target block. “Block located at the corner of the target block” may have the same meaning as “block adjacent to the corner of the target block.” “Blocks located at the corners of the target block” may be included in “blocks adjacent to the target block.”

For example, spatial candidates include a reconstructed block located to the left of the target block, a reconstructed block located at the top of the target block, a reconstructed block located at the lower left corner of the target block, and a reconstructed block located at the upper right corner of the target block. It may be a reconstructed block or a reconstructed block located in the upper left corner of the target block.

Each of the encoding device 100 and the decoding device 200 can identify a block that exists at a location spatially corresponding to the target block within a coll picture. The location of the target block in the target picture and the location of the identified block in the call picture may correspond to each other.

Each of the encoding device 100 and the decoding device 200 may determine a col block existing at a predefined relative position with respect to the identified block as a temporal candidate. The predefined relative position may be a position inside and/or outside the identified block.

For example, a call block may include a first call block and a second call block. When the coordinates of the identified block are (xP, yP) and the size of the identified block is (nPSW, nPSH), the first call block may be a block located at the coordinates (xP + nPSW, yP + nPSH). The second call block may be a block located at the coordinates (xP + (nPSW >> 1), yP + (nPSH >> 1)). The second call block can be selectively used when the first call block is unavailable.

The motion vector of the target block may be determined based on the motion vector of the call block. Each of the encoding device 100 and the decoding device 200 can scale the motion vector of a call block. The scaled motion vector of the call block can be used as the motion vector of the target block. Additionally, the motion vector of the motion information of the temporal candidate stored in the list may be a scaled motion vector.

The ratio of the motion vector of the target block and the motion vector of the call block may be equal to the ratio of the first temporal distance and the second temporal distance. The first temporal distance may be the distance between the reference picture of the target block and the target picture. The second temporal distance may be the distance between the reference picture and the call picture of the call block.

The method of deriving motion information may vary depending on the inter prediction mode of the target block. For example, inter prediction modes applied for inter prediction include Advanced Motion Vector Predictor (AMVP) mode, merge mode and skip mode, merge mode with motion vector difference, There may be subblock merge mode, triangulation mode, inter-intra combined prediction mode, affine inter mode, and current picture reference mode. Merge mode may also be referred to as motion merge mode. Below, each of the modes is explained in detail.

1) AMVP mode

When AMVP mode is used, the encoding device 100 can search for similar blocks in the neighbors of the target block. The encoding device 100 may obtain a prediction block by performing prediction on the target block using motion information of the searched similar block. The encoding device 100 may encode a residual block that is the difference between the target block and the prediction block.

1-1) Creation of a predicted motion vector candidate list

When the AMVP mode is used as the prediction mode, each of the encoding device 100 and the decoding device 200 can generate a prediction motion vector candidate list using the motion vector of the spatial candidate, the motion vector of the temporal candidate, and the zero vector. there is. The predicted motion vector candidate list may include one or more predicted motion vector candidates. At least one of the motion vector of the spatial candidate, the motion vector of the temporal candidate, and the zero vector may be determined and used as the predicted motion vector candidate.

Hereinafter, the terms “predicted motion vector (candidate)” and “motion vector (candidate)” may be used with the same meaning and may be used interchangeably.

Hereinafter, the terms “predicted motion vector candidate” and “AMVP candidate” may be used with the same meaning and may be used interchangeably.

Hereinafter, the terms “predicted motion vector candidate list” and “AMVP candidate list” may be used with the same meaning and may be used interchangeably.

Spatial candidates may include reconstructed spatial neighboring blocks. In other words, the motion vector of the reconstructed neighboring block may be referred to as a spatial prediction motion vector candidate.

Temporal candidates may include call blocks and blocks adjacent to call blocks. That is, the motion vector of a call block or a motion vector of a block adjacent to a call block may be referred to as a temporal prediction motion vector candidate.

The zero vector may be a (0, 0) motion vector.

The predicted motion vector candidate may be a motion vector predictor for predicting a motion vector. Additionally, in the encoding device 100, a predicted motion vector candidate may be a motion vector initial search position.

1-2) Search for motion vector using predicted motion vector candidate list

The encoding device 100 may use the predicted motion vector candidate list to determine a motion vector to be used for encoding the target block within the search range. Additionally, the encoding device 100 may determine a prediction motion vector candidate to be used as the prediction motion vector of the target block among the prediction motion vector candidates in the prediction motion vector candidate list.

The motion vector to be used for encoding the target block may be a motion vector that can be encoded at minimal cost.

Additionally, the encoding device 100 may determine whether to use the AMVP mode when encoding the target block.

1-3) Transmission of inter prediction information

The encoding device 100 may generate a bitstream including inter prediction information required for inter prediction. The decoding device 200 may perform inter prediction on the target block using inter prediction information of the bitstream.

Inter prediction information includes 1) mode information indicating whether AMVP mode is used, 2) prediction motion vector index, 3) motion vector difference (MVD), 4) reference direction, and 5) reference picture index. can do.

Hereinafter, the terms “predicted motion vector index” and “AMVP index” may be used with the same meaning and may be used interchangeably.

Additionally, inter prediction information may include a residual signal.

When the mode information indicates that AMVP mode is used, the decoding device 200 may obtain a predicted motion vector index, motion vector difference, reference direction, and reference picture index from the bitstream through entropy decoding.

The prediction motion vector index may indicate a prediction motion vector candidate used for prediction of the target block among prediction motion vector candidates included in the prediction motion vector candidate list.

1-4) Inter prediction in AMVP mode using inter prediction information

The decoding apparatus 200 may derive a predicted motion vector candidate using the predicted motion vector candidate list and determine motion information of the target block based on the derived predicted motion vector candidate.

The decoding apparatus 200 may use the predicted motion vector index to determine a motion vector candidate for the target block among the predicted motion vector candidates included in the predicted motion vector candidate list. The decoding apparatus 200 may select the prediction motion vector candidate indicated by the prediction motion vector index from among the prediction motion vector candidates included in the prediction motion vector candidate list as the prediction motion vector of the target block.

The encoding device 100 may generate an entropy-encoded predicted motion vector index by applying entropy coding to the predicted motion vector index, and generate a bitstream including the entropy-encoded predicted motion vector index. The entropy-encoded predicted motion vector index may be signaled from the encoding device 100 to the decoding device 200 through a bitstream. The decoding device 200 can extract an entropy-encoded predicted motion vector index from a bitstream, and obtain the predicted motion vector index by applying entropy decoding to the entropy-encoded predicted motion vector index.

The motion vector actually used for inter prediction of the target block may not match the prediction motion vector. MVD may be used to represent the motion vector that will actually be used for inter prediction of the target block and the difference between the prediction motion vectors. The encoding device 100 may derive a prediction motion vector similar to the motion vector that will actually be used for inter prediction of the target block in order to use the MVD of the smallest size possible.

MVD may be the difference between the motion vector of the target block and the predicted motion vector. The encoding device 100 can calculate the MVD and generate an entropy-encoded MVD by applying entropy encoding to the MVD. The encoding device 100 may generate a bitstream including entropy-encoded MDV.

MVD may be transmitted from the encoding device 100 to the decoding device 200 through a bitstream. The decoding device 200 can extract the entropy-encoded MVD from the bitstream and obtain the MVD by applying entropy decoding to the entropy-encoded MVD.

The decoding device 200 can derive the motion vector of the target block by combining the MVD and the predicted motion vector. In other words, the motion vector of the target block derived from the decoding device 200 may be the sum of the MVD and the motion vector candidate.

Additionally, the encoding device 100 can generate entropy-encoded MVD resolution information by applying entropy encoding to the calculated MVD resolution information, and can generate a bitstream including the entropy-encoded MVD resolution information. The decoding device 200 can extract entropy-encoded MVD resolution information from the bitstream and obtain MVD resolution information by applying entropy decoding to the entropy-encoded MVD resolution information. The decoding device 200 can adjust the resolution of the MVD using the MVD resolution information.

Meanwhile, the encoding device 100 may calculate the MVD based on the affine model. The decoding device 200 may derive an affine control motion vector of the target block through the sum of the MVD and affine control motion vector candidates, and may derive a motion vector for the sub-block using the affine control motion vector. there is.

The reference direction may indicate a reference picture list used for prediction of the target block. For example, the reference direction may point to one of the reference picture list L0 and the reference picture list L1.

The reference direction only indicates a reference picture list used for prediction of the target block, and may not indicate that the directions of reference pictures are limited to the forward direction or backward direction. That is, each of the reference picture list L0 and the reference picture list L1 may include forward and/or reverse pictures.

The fact that the reference direction is uni-directional may mean that one reference picture list is used. Bi-directional reference direction may mean that two reference picture lists are used. That is, the reference direction may indicate that only the reference picture list L0 is used, that only the reference picture list L1 is used, and one of the two reference picture lists.

The reference picture index may indicate a reference picture used for prediction of the target block among reference pictures in the reference picture list. The encoding device 100 can generate an entropy-coded reference picture index by applying entropy coding to the reference picture index and generate a bitstream including the entropy-coded reference picture index. The entropy-coded reference picture index may be signaled from the encoding device 100 to the decoding device 200 through a bitstream. The decoding device 200 can extract an entropy-coded reference picture index from a bitstream and obtain the reference picture index by applying entropy decoding to the entropy-coded reference picture index.

When two reference picture lists are used for prediction of the target block. One reference picture index and one motion vector can be used for each reference picture list. Additionally, when two reference picture lists are used for prediction of the target block, two prediction blocks may be specified for the target block. For example, a (final) prediction block of the target block may be generated through an average or weighted sum of two prediction blocks for the target block.

The motion vector of the target block can be derived by the predicted motion vector index, MVD, reference direction, and reference picture index.

The decoding device 200 may generate a prediction block for the target block based on the derived motion vector and reference picture index. For example, the prediction block may be a reference block pointed to by the derived motion vector in the reference picture indicated by the reference picture index.

By encoding the predicted motion vector index and MVD rather than encoding the motion vector of the target block itself, the amount of bits transmitted from the encoding device 100 to the decoding device 200 can be reduced and coding efficiency can be improved.

The motion information of the reconstructed neighboring block may be used for the target block. In a specific inter prediction mode, the encoding device 100 may not separately encode the motion information itself for the target block. The motion information of the target block is not encoded, and other information that can derive the motion information of the target block through the motion information of the reconstructed neighboring block may be encoded instead. As other information is encoded instead, the amount of bits transmitted to the decoding device 200 can be reduced and coding efficiency can be improved.

For example, as an inter prediction mode in which the motion information of the target block is not directly encoded, there may be a skip mode and/or a merge mode. At this time, the encoding device 100 and the decoding device 200 may use an identifier and/or index that indicates which unit's motion information among the reconstructed neighboring units is used as the motion information of the target unit.

2) Merge Mode

As a method of deriving the movement information of the target block, there is a merge. Merge may mean merging movements of multiple blocks. Merge may mean applying the movement information of one block to other blocks as well. In other words, the merge mode may mean a mode in which the motion information of the target block is derived from the motion information of the neighboring block.

When the merge mode is used, the encoding device 100 may perform prediction on the motion information of the target block using motion information of the spatial candidate and/or motion information of the temporal candidate. Spatial candidates may include reconstructed spatial neighboring blocks that are spatially adjacent to the target block. Spatial neighboring blocks may include left neighboring blocks and top neighboring blocks. Temporal candidates may include call blocks. The terms “spatial candidate” and “spatial merge candidate” may be used interchangeably and may be used interchangeably. The terms “temporal candidate” and “temporal merge candidate” may be used interchangeably and may be used interchangeably.

The encoding device 100 may obtain a prediction block through prediction. The encoding device 100 may encode a residual block that is the difference between the target block and the prediction block.

2-1) Creation of merge candidate list

When the merge mode is used, each of the encoding device 100 and the decoding device 200 may generate a merge candidate list using motion information of the spatial candidate and/or motion information of the temporal candidate. Motion information may include 1) a motion vector, 2) a reference picture index, and 3) a reference direction. The reference direction can be unidirectional or bidirectional. The reference direction may mean an inter prediction indicator.

The merge candidate list may include merge candidates. The merge candidate may be motion information. In other words, the merge candidate list may be a list in which motion information is stored.

Merge candidates may be motion information such as temporal candidates and/or spatial candidates. In other words, the merge candidate list may include motion information such as temporal candidates and/or spatial candidates.

Additionally, the merge candidate list may include a new merge candidate created by combining merge candidates that already exist in the merge candidate list. In other words, the merge candidate list may include new motion information generated by combining motion information that already exists in the merge candidate list.

Additionally, the merge candidate list may include history-based merge candidates. A history-based merge candidate may be motion information of a block that was encoded and/or decoded before the target block.

Additionally, the merge candidate list may include a merge candidate based on the average of two merge candidates.

Merge candidates may be specified modes that derive inter prediction information. A merge candidate may be information indicating a specified mode that derives inter prediction information. Inter prediction information of the target block can be derived according to the specified mode indicated by the merge candidate. At this time, the specified mode may include a process of deriving a series of inter prediction information. This specified mode may be an inter prediction information derivation mode or a motion information derivation mode.

Inter prediction information of the target block may be derived according to the mode indicated by the merge candidate selected by the merge index among the merge candidates in the merge candidate list.

For example, the motion information derivation modes in the merge candidate list may be at least one of 1) a motion information derivation mode on a sub-block basis and 2) an affine motion information derivation mode.

Additionally, the merge candidate list may include motion information of the zero vector. Zero vectors may also be referred to as zero merge candidates.

In other words, the motion information in the merge candidate list is: 1) motion information of the spatial candidate, 2) motion information of the temporal candidate, 3) motion information generated by a combination of motion information already existing in the merge candidate list, and 4) zero vector. It can be at least one of:

Motion information may include 1) a motion vector, 2) a reference picture index, and 3) a reference direction. The reference direction may also be referred to as an inter prediction indicator. The reference direction can be unidirectional or bidirectional. A unidirectional reference direction may represent L0 prediction or L1 prediction.

The merge candidate list can be created before prediction by merge mode is performed.

The number of merge candidates in the merge candidate list may be predefined. The encoding device 100 and the decoding device 200 may add merge candidates to the merge candidate list according to a predefined method and a predefined rank so that the merge candidate list has a predefined number of merge candidates. Through a predefined method and a predefined ranking, the merge candidate list of the encoding device 100 and the merge candidate list of the decoding device 200 may be the same.

Merge can be applied on a CU or PU basis. When merging is performed on a CU or PU basis, the encoding device 100 may transmit a bitstream containing predefined information to the decoding device 200. For example, predefined information includes 1) information indicating whether to perform a merge for each block partition, 2) which block to merge with among blocks that are spatial candidates and/or temporal candidates for the target block. It may include information about whether

2-2) Search for motion vector using merge candidate list

The encoding device 100 may determine a merge candidate to be used for encoding the target block. For example, the encoding device 100 may perform predictions on a target block using merge candidates from a merge candidate list and generate residual blocks for the merge candidates. The encoding device 100 may use a merge candidate that requires the minimum cost in encoding the prediction and residual blocks to encode the target block.

Additionally, the encoding device 100 may determine whether to use merge mode when encoding the target block.

2-3) Transmission of inter prediction information

The encoding device 100 may generate a bitstream including inter prediction information required for inter prediction. The encoding device 100 may generate entropy-encoded inter prediction information by performing entropy encoding on the inter prediction information, and may transmit a bitstream including the entropy-encoded inter prediction information to the decoding device 200. Through the bitstream, entropy-encoded inter prediction information may be signaled from the encoding device 100 to the decoding device 200. The decoding device 200 can extract entropy-encoded inter prediction information from a bitstream and obtain inter-prediction information by performing entropy decoding on the entropy-encoded inter prediction information.

The decoding device 200 may perform inter prediction on the target block using inter prediction information of the bitstream.

Inter prediction information may include 1) mode information indicating whether to use merge mode, 2) merge index, and 3) correction information.

Additionally, inter prediction information may include a residual signal.

The decoding device 200 can obtain a merge index from the bitstream only when the mode information indicates that the merge mode is used.

Mode information may be a merge flag. The unit of mode information may be a block. Information about the block may include mode information, and the mode information may indicate whether merge mode is applied to the block.

The merge index may indicate a merge candidate used to predict the target block among the merge candidates included in the merge candidate list. Alternatively, the merge index may indicate with which block among neighboring blocks spatially or temporally adjacent to the target block the merge is performed.

The encoding device 100 may select a merge candidate with the highest coding performance among the merge candidates included in the merge candidate list, and set the value of the merge index to indicate the selected merge candidate.

Correction information may be information used to correct a motion vector. The encoding device 100 can generate correction information. The decoding device 200 may correct the motion vector of the merge candidate selected by the merge index based on the correction information.

Correction information may include at least one of information indicating whether correction is made, correction direction information, and correction size information. The prediction mode that corrects the motion vector based on the signaled correction information may be called a merge mode with motion vector difference.

2-4) Inter prediction in merge mode using inter prediction information

The decoding device 200 may perform prediction on the target block using the merge candidate indicated by the merge index among the merge candidates included in the merge candidate list.

The motion vector of the target block can be specified by the motion vector of the merge candidate indicated by the merge index, the reference picture index, and the reference direction.

3) Skip mode

Skip mode may be a mode in which motion information of a spatial candidate or motion information of a temporal candidate is applied to the target block as is. Additionally, the skip mode may be a mode that does not use a residual signal. That is, when skip mode is used, the reconstructed block may be identical to the prediction block.

The difference between merge mode and skip mode may be whether or not residual signals are transmitted or used. That is to say, skip mode may be similar to merge mode except that no residual signals are transmitted or used.

When skip mode is used, the encoding device 100 sends information indicating which block's motion information among spatial candidate or temporal candidate blocks is used as motion information of the target block to the decoding device 200 through a bitstream. Can be transmitted. The encoding device 100 can generate entropy-coded information by performing entropy encoding on such information, and can signal the entropy-coded information to the decoding device 200 through a bitstream. The decoding device 200 can extract entropy-encoded information from a bitstream and obtain information by performing entropy decoding on the entropy-encoded information.

Additionally, when skip mode is used, the encoding device 100 may not transmit other syntax element information, such as MVD, to the decoding device 200. For example, when skip mode is used, the encoding device 100 may not signal syntax elements related to at least one of the MVD, the coded block flag, and the transform coefficient level to the decoding device 200.

3-1) Creation of merge candidate list

Skip mode can also use the merge candidate list. That is, the merge candidate list can be used in both merge mode and skip mode. In this respect, the merge candidate list may be named “skip candidate list” or “merge/skip candidate list.”

Alternatively, skip mode may use a separate candidate list than merge mode. In this case, in the description below, the merge candidate list and merge candidate may be replaced with the skip candidate list and skip candidate, respectively.

The merge candidate list can be created before prediction by skip mode is performed.

3-2) Search for motion vector using merge candidate list

The encoding device 100 may determine a merge candidate to be used for encoding the target block. For example, the encoding device 100 may perform predictions on the target block using merge candidates from the merge candidate list. The encoding device 100 may use a merge candidate that requires the minimum cost in prediction to encode the target block.

Additionally, the encoding device 100 may determine whether to use skip mode when encoding the target block.

3-3) Transmission of inter prediction information

The encoding device 100 may generate a bitstream including inter prediction information required for inter prediction. The decoding device 200 may perform inter prediction on the target block using inter prediction information of the bitstream.

Inter prediction information may include 1) mode information indicating whether skip mode is used, and 2) skip index.

The skip index may be the same as the merge index described above.

When skip mode is used, the target block can be encoded without a residual signal. Inter prediction information may not include residual signals. Alternatively, the bitstream may not include a residual signal.

The decoding device 200 can obtain a skip index from the bitstream only when the mode information indicates that skip mode is used. As described above, the merge index and skip index may be the same. The decoding device 200 can obtain a skip index from the bitstream only when the mode information indicates that merge mode or skip mode is used.

The skip index may indicate a merge candidate used to predict the target block among the merge candidates included in the merge candidate list.

3-4) Inter prediction in skip mode using inter prediction information

The decoding device 200 may perform prediction on the target block using the merge candidate indicated by the skip index among the merge candidates included in the merge candidate list.

The motion vector of the target block can be specified by the motion vector of the merge candidate indicated by the skip index, the reference picture index, and the reference direction.

4) Current picture reference mode

The current picture reference mode may refer to a prediction mode that uses a pre-reconstructed area within the target picture to which the target block belongs.

A motion vector may be used to specify a pre-reconstructed area. Whether the target block is encoded in the current picture reference mode can be determined using the reference picture index of the target block.

A flag or index indicating whether the target block is a block encoded in the current picture reference mode may be signaled from the encoding device 100 to the decoding device 200. Alternatively, whether the target block is a block encoded in the current picture reference mode may be inferred through the reference picture index of the target block.

When the target block is encoded in the current picture reference mode, the target picture may exist at a fixed position or a random position within the reference picture list for the target block.

For example, the fixed position may be a position where the reference picture index value is 0 or the very last position.

If the target picture exists at a random position in the reference picture list, a separate reference picture index indicating this random position may be signaled from the coding device 100 to the decoding device 200.

5) Subblock merge mode

Subblock merge mode may refer to a mode that derives motion information for a subblock of a CU.

When the sub-block merge mode is applied, motion information of the call sub-block of the target sub-block in the reference image (i.e., sub-block based temporal merge candidate) and/or affine control point motion vector A subblock merge candidate list may be created using a merge candidate (affine control point motion vector merge candidate).

6) Triangle partition mode

In triangulation mode, divided target blocks can be created by dividing the target block diagonally. For each divided target block, motion information of each divided target block may be derived, and prediction samples for each divided target block may be derived using the derived motion information. The prediction sample of the target block may be derived through a weighted sum of the prediction samples of the divided target blocks.

7) Inter-intra combined prediction mode

The inter-intra combined prediction mode may be a mode in which a prediction sample of the target block is derived using a weighted sum of prediction samples generated by inter prediction and prediction samples generated by intra prediction.

In the above-described modes, the decoding device 200 can perform its own correction on the derived motion information. For example, the decoding device 200 may search a specified area based on the reference block indicated by the derived motion information and search for motion information with the minimum sum of absolute differences (SAD). And, the searched motion information can be derived as corrected motion information.

In the above-described modes, the decoding device 200 may perform compensation for prediction samples derived through inter prediction using optical flow.

In the above-described AMVP mode, merge mode, skip mode, etc., motion information to be used for prediction of the target block among motion information in the list can be specified through an index to the list.

To improve coding efficiency, the encoding device 100 may signal only the index of the element that causes the minimum cost in inter prediction of the target block among the elements of the list. The encoding device 100 can encode an index and signal the encoded index.

Accordingly, the above-described lists (i.e., the predicted motion vector candidate list and the merge candidate list) may have to be derived in the same manner based on the same data in the encoding device 100 and the decoding device 200. Here, the same data may include a reconstructed picture and a reconstructed block. Additionally, in order to specify elements by index, the order of elements within the list may need to be constant.

Figure 10 shows spatial candidates according to an example.

In Figure 10, the locations of spatial candidates are shown.

The large block in the middle may represent the target block. Five small blocks may represent spatial candidates.

The coordinates of the target block may be (xP, yP), and the size of the target block may be (nPSW, nPSH).

Spatial candidate A ₀ may be a block adjacent to the lower left corner of the target block. A ₀ may be a block occupying a pixel with coordinates (xP - 1, yP + nPSH).

Spatial candidate A ₁ may be a block adjacent to the left of the target block. A ₁ may be the lowest block among blocks adjacent to the left of the target block. Alternatively, A ₁ may be a block adjacent to the top of A ₀ . A ₁ may be a block occupying a pixel with coordinates (xP - 1, yP + nPSH - 1).

Spatial candidate B ₀ may be a block adjacent to the upper right corner of the target block. B ₀ may be a block occupying a pixel with coordinates (xP + nPSW, yP - 1).

Spatial candidate B ₁ may be a block adjacent to the top of the target block. B ₁ may be the rightmost block among blocks adjacent to the top of the target block. Alternatively, B ₁ may be a block adjacent to the left of B ₀ . B ₁ may be a block occupying a pixel with coordinates (xP + nPSW - 1, yP - 1).

Spatial candidate B ₂ may be a block adjacent to the upper left corner of the target block. B ₂ may be a block occupying a pixel with coordinates (xP - 1, yP - 1).

Determination of availability of spatial and temporal candidates

In order to include the motion information of the spatial candidate or the motion information of the temporal candidate in the list, it must be determined whether the motion information of the spatial candidate or the motion information of the temporal candidate is available.

Hereinafter, candidate blocks may include spatial candidates and temporal candidates.

For example, the above determination can be made by sequentially applying steps 1) to 4) below.

Step 1) If the PU containing the candidate block is outside the boundary of the picture, the availability of the candidate block may be set to false. “Availability is set to false” can mean the same as “set to unavailability.”

Step 2) If the PU containing the candidate block is outside the boundary of the slice, the availability of the candidate block may be set to false. If the target block and the candidate block are located in different slices, the availability of the candidate block may be set to false.

Step 3) If the PU containing the candidate block is outside the boundary of the tile, the availability of the candidate block may be set to false. If the target block and the candidate block are located within different tiles, the availability of the candidate block may be set to false.

Step 4) If the prediction mode of the PU including the candidate block is intra prediction mode, the availability of the candidate block may be set to false. If the PU containing the candidate block does not use inter prediction, the availability of the candidate block may be set to false.

Figure 11 shows the order of adding motion information of spatial candidates to a merge list according to an example.

As shown in FIG. 11, when adding motion information of spatial candidates to the merge list, the order of A ₁ , B ₁ , B ₀ , A ₀ and B ₂ can be used. That is, motion information of available spatial candidates may be added to the merge list in the following order: A ₁ , B ₁ , B ₀ , A ₀ , and B ₂ .

Merge list derivation method in merge mode and skip mode

As described above, the maximum number of merge candidates in the merge list can be set. The set maximum number is indicated as N. The set number may be transmitted from the encoding device 100 to the decoding device 200. The slice header of a slice may include N. In other words, the maximum number of merge candidates in the merge list for the target block of the slice can be set by the slice header. For example, by default, the value of N may be 5.

Motion information (i.e., merge candidate) can be added to the merge list in the order of steps 1) to 4) below.

Step 1) Among the spatial candidates, available spatial candidates can be added to the merge list. Motion information of available spatial candidates can be added to the merge list in the order shown in FIG. 11. At this time, if the motion information of the available spatial candidate overlaps with other motion information that already exists in the merge list, the motion information may not be added to the merge list. Checking whether there is overlap with other motion information present in the list can be outlined as a “redundancy check.”

There may be a maximum of N pieces of added motion information.

Step 2) If the number of motion information items in the merge list is smaller than N and a temporal candidate is available, the motion information of the temporal candidate may be added to the merge list. At this time, if the motion information of the available temporal candidate overlaps with other motion information that already exists in the merge list, the motion information may not be added to the merge list.

Step 3) If the number of motion information in the merge list is smaller than N and the type of target slice is "B", the combined motion information generated by combined bi-prediction will be added to the merge list. You can.

The target slice may be a slice containing the target block.

The combined motion information may be a combination of L0 motion information and L1 motion information. L0 motion information may be motion information that refers only to the reference picture list L0. L1 motion information may be motion information that refers only to the reference picture list L1.

Within the merge list, there may be more than one L0 motion information. Additionally, within the merge list, there may be more than one L1 motion information.

There may be more than one piece of combined motion information. In generating the combined motion information, which L0 motion information and which L1 motion information to use among one or more L0 motion information and one or more L1 motion information may be predefined. One or more combined motion information may be generated in a predefined order by combined bidirectional prediction using pairs of different motion information in the merge list. One of the pairs of different motion information may be L0 motion information and the other may be L1 motion information.

For example, the combined motion information added with highest priority may be a combination of L0 motion information with a merge index of 0 and L1 motion information with a merge index of 1. If motion information with a merge index of 0 is not L0 motion information, or motion information with a merge index of 1 is not L1 motion information, the above combined motion information may not be generated and added. The motion information added next may be a combination of L0 motion information with a merge index of 1 and L1 motion information with a merge index of 0. The specific combination below may follow other combinations in the video encoding/decoding field.

At this time, if the combined motion information overlaps with other motion information that already exists in the merge list, the combined motion information may not be added to the merge list.

Step 4) If the number of motion information items in the merge list is smaller than N, zero vector motion information may be added to the merge list.

Zero vector motion information may be motion information in which the motion vector is a zero vector.

There may be one or more zero vector motion information. Reference picture indices of one or more pieces of zero vector motion information may be different from each other. For example, the value of the reference picture index of the first zero vector motion information may be 0. The value of the reference picture index of the second zero vector motion information may be 1.

The number of zero vector motion information may be equal to the number of reference pictures in the reference picture list.

The reference direction of zero vector motion information may be bidirectional. Both motion vectors may be zero vectors. The number of zero vector motion information may be the smaller of the number of reference pictures in the reference picture list L0 and the number of reference pictures in the reference picture list L1. Alternatively, if the number of reference pictures in the reference picture list L0 and the number of reference pictures in the reference picture list L1 are different from each other, a unidirectional reference direction may be used for a reference picture index that can be applied to only one reference picture list.

The encoding device 100 and/or the decoding device 200 may sequentially add zero vector motion information to the merge list while changing the reference picture index.

If the zero vector motion information overlaps with other motion information that already exists in the merge list, the zero vector motion information may not be added to the merge list.

The order of steps 1) to 4) described above is merely exemplary, and the order between steps may be changed. Additionally, some of the steps may be omitted depending on predefined conditions.

Method for deriving a predicted motion vector candidate list in AMVP mode

The maximum number of prediction motion vector candidates in the prediction motion vector candidate list may be predefined. The predefined maximum number is denoted by N. For example, the predefined maximum number may be 2.

Motion information (i.e., predicted motion vector candidate) may be added to the predicted motion vector candidate list in the order of steps 1) to 3) below.

Step 1) Available spatial candidates among spatial candidates may be added to the predicted motion vector candidate list. Spatial candidates may include a first spatial candidate and a second spatial candidate.

The first spatial candidate may be one of A ₀ , A ₁ , scaled A ₀ and scaled A ₁ . The second spatial candidate may be one of B ₀ , B ₁ , B ₂ , scaled B ₀ , scaled B ₁ and scaled B ₂ .

Motion information of available spatial candidates may be added to the predicted motion vector candidate list in the order of the first spatial candidate and the second spatial candidate. At this time, if the motion information of the available spatial candidate overlaps with other motion information that already exists in the prediction motion vector candidate list, the motion information may not be added to the prediction motion vector candidate list. In other words, when the value of N is 2, if the motion information of the second spatial candidate is the same as the motion information of the first spatial candidate, the motion information of the second spatial candidate may not be added to the prediction motion vector candidate list.

There may be a maximum of N pieces of added motion information.

Step 2) If the number of motion information items in the predicted motion vector candidate list is smaller than N and a temporal candidate is available, the motion information of the temporal candidate may be added to the predicted motion vector candidate list. At this time, if the motion information of the available temporal candidate overlaps with other motion information that already exists in the predicted motion vector candidate list, the motion information may not be added to the predicted motion vector candidate list.

Step 3) If the number of motion information pieces in the predicted motion vector candidate list is smaller than N, zero vector motion information may be added to the predicted motion vector candidate list.

There may be one or more zero vector motion information. Reference picture indices of one or more pieces of zero vector motion information may be different from each other.

The encoding device 100 and/or the decoding device 200 may sequentially add zero vector motion information to the prediction motion vector candidate list while changing the reference picture index.

If the zero vector motion information overlaps with other motion information that already exists in the prediction motion vector candidate list, the zero vector motion information may not be added to the prediction motion vector candidate list.

The description of the zero vector motion information described above for the merge list can also be applied to the zero vector motion information. Redundant descriptions are omitted.

The order of steps 1) to 3) described above is merely exemplary, and the order between steps may be changed. Additionally, some of the steps may be omitted depending on predefined conditions.

Figure 12 explains the process of conversion and quantization according to an example.

As shown in FIG. 12, a quantized level can be generated by performing a conversion and/or quantization process on the residual signal.

The residual signal can be generated as the difference between the original block and the prediction block. Here, the prediction block may be a block generated by intra prediction or inter prediction.

The residual signal can be converted to the frequency domain through a transformation process that is part of the quantization process.

Transformation kernels used for transformation may include various DCT kernels such as Discrete Cosine Transform (DCT) type 2 (DCT-II) and Discrete Sine Transform (DST) kernels. .

These transform kernels can perform a separable transform or a 2Dimensional (2D) non-separable transform on the residual signal. The separable transformation may be a transformation that performs one-dimensional (1D) transformation on the residual signal in each of the horizontal and vertical directions.

DCT types and DST types adaptively used for 1D conversion may include DCT-V, DCT-VIII, DST-I, and DST-VII in addition to DCT-II, as shown in Table 3 and Table 4 below, respectively. there is.

[Table 3]

[Table 4]

As shown in Tables 3 and 4, a transform set can be used to derive the DCT type or DST type to be used for transformation. Each transformation set may include multiple transformation candidates. Each transformation candidate may be a DCT type or a DST type.

Table 5 below shows an example of a transform set applied to the horizontal direction and a transform set applied to the vertical direction according to the intra prediction mode.

[Table 5]

In Table 5, the number of the vertical transformation set and the number of the horizontal transformation set applied to the horizontal direction of the residual signal according to the intra prediction mode of the target block are displayed.

As illustrated in Table 5, transformation sets applied to the horizontal and vertical directions may be predefined according to the intra prediction mode of the target block. The encoding device 100 may perform transformation and inverse transformation on the residual signal using the transformation included in the transformation set corresponding to the intra prediction mode of the target block. Additionally, the decoding apparatus 200 may perform inverse transformation on the residual signal using a transformation included in a transformation set corresponding to the intra prediction mode of the target block.

For these transformations and inverse transformations, the set of transformations applied to the residual signal may be determined as illustrated in Tables 3, 4, and 5, and may be unsignaled. Transformation instruction information may be signaled from the encoding device 100 to the decoding device 200. Transformation instruction information may be information indicating which transform candidate is used among a plurality of transform candidates included in a transform set applied to the residual signal.

For example, when the size of the target block is 64x64 or less, transform sets each having three transforms may be configured according to the intra prediction mode. The optimal transformation method can be selected among a total of 9 multiple transformation methods resulting from a combination of three transformations in the horizontal direction and three transformations in the vertical direction. Coding efficiency can be improved by encoding and/or decoding the residual signal using this optimal conversion method.

At this time, for at least one of the vertical transformation and the horizontal transformation, information about which transformation among the transformations belonging to the transformation set was used may be entropy encoded and/or decoded. Truncated unary binarization may be used to encode and/or decode this information.

The method using various transforms as described above can be applied to a residual signal generated by intra prediction or inter prediction.

Transformation may include at least one of primary transformation and secondary transformation. A transform coefficient can be generated by performing a first-order transform on the residual signal, and a second-order transform coefficient can be generated by performing a second-order transform on the transform coefficient.

A primary transformation may be named primary. Additionally, the first-order transform may be named Adaptive Multiple Transform (AMT). AMT may mean that different transformations are applied to each of the 1D directions (i.e., vertical and horizontal directions) as described above.

The secondary transformation may be a transformation to improve the energy concentration of the transformation coefficient generated by the primary transformation. Secondary transformations, like primary transformations, can be either separable transformations or non-separable transformations. The non-separable transform may be a Non-Separable Secondary Transform (NSST).

Primary transformation may be performed using at least one of a plurality of predefined transformation methods. For example, a plurality of predefined transformation methods include Discrete Cosine Transform (DCT), Discrete Sine Transform (DST), and Karhunen-Loeve Transform (KLT)-based transformation, etc. It can be included.

Additionally, the first-order transformation may be a transformation with various transformation types depending on the kernel function that defines DCT or DST.

For example, the transformation type is 1) prediction mode of the target block (e.g., one of intra prediction and inter prediction), 2) size of the target block, 3) shape of the target block, 4) intra prediction mode of the target block. , 5) a component of the target block (e.g., one of the luma component and a chroma component), and 6) the partition type applied to the target block (e.g., Quad Tree (QT), Binary Tree (BT) ) and one of a Ternary Tree (TT).

For example, the first-order transformation includes transformations such as DCT-2, DCT-5, DCT-7, DST-7, DST-1, DST-8, and DCT-8 according to the transformation kernels shown in Table 6 below. can do. Table 6 illustrates various transform types and transform kernel functions for multiple transform selection (MTS).

MTS may mean that a combination of one or more DCT and/or DST transformation kernels is selected to transform the residual signal in the horizontal and/or vertical directions.

[Table 6]

In Table 6, i and j may be integer values between 0 and N-1.

A secondary transform may be performed on the transformation coefficient generated by performing the primary transformation.

As with first-order transformations, a set of transformations can be defined for second-order transformations. Methods for deriving and/or determining a set of transformations such as those described above can be applied to secondary transformations as well as primary transformations.

Primary transformation and secondary transformation can be determined for a specified target.

For example, a first-order transform and a second-order transform may be applied to one or more signal components of a luma component and a chroma component. Whether to apply the first transform and/or the second transform may be determined according to at least one of coding parameters for the target block and/or the neighboring block. For example, whether to apply primary transformation and/or secondary transformation may be determined by the size and/or shape of the target block.

In the encoding device 100 and the decoding device 200, conversion information indicating the conversion method to be used for the target can be derived by using specified information.

For example, the transformation information may include an index of the transformation to be used for primary transformation and/or secondary transformation. Alternatively, the transformation information may indicate that the primary transformation and/or secondary transformation is not used.

For example, when the target of the primary transformation and secondary transformation is the target block, the transformation method(s) applied to the primary transformation and/or secondary transformation indicated by the transformation information is applied to the target block and/or neighboring blocks. It may be determined according to at least one of the coding parameters for.

Alternatively, conversion information indicating a conversion method for a specified target may be signaled from the encoding device 100 to the decoding device 200.

For example, for one CU, whether the primary transform is used, an index indicating the primary transform, whether the secondary transform is used, and an index indicating the secondary transform, etc. can be derived as transformation information in the decoding device 200. there is. Alternatively, for one CU, transformation information indicating whether to use the primary transformation, an index indicating the primary transformation, whether to use the secondary transformation, and an index indicating the secondary transformation may be signaled.

A quantized transform coefficient (i.e., a quantized level) may be generated by performing quantization on a result or a residual signal generated by performing a first-order transform and/or a second-order transform.

13 shows diagonal scanning according to an example.

14 shows horizontal scanning according to an example.

15 shows vertical scanning according to one example.

The quantized transform coefficients may be scanned according to at least one of (up-right) diagonal scanning, vertical scanning, and horizontal scanning, according to at least one of intra prediction mode, block size, and block type. A block may be a transformation unit.

Each scanning can start at a specified starting point and end at a specified ending point.

For example, by scanning the coefficients of a block using the diagonal scanning of FIG. 13, the quantized transform coefficients can be changed into a one-dimensional vector form. Alternatively, the horizontal scanning of FIG. 14 or the vertical scanning of FIG. 15 may be used instead of diagonal scanning, depending on the size of the block and/or the intra prediction mode.

Vertical scanning may be scanning two-dimensional block-shaped coefficients in a column direction. Horizontal scanning may be scanning two-dimensional block-shaped coefficients in the row direction.

That is, depending on the size of the block and/or the inter prediction mode, it may be determined which scanning among diagonal scanning, vertical scanning, and horizontal scanning will be used.

As shown in FIGS. 13, 14, and 15, the quantized transform coefficients can be scanned along the diagonal, horizontal, or vertical directions.

Quantized transform coefficients can be expressed in block form. A block may include multiple sub-blocks. Each subblock can be defined according to the minimum block size or minimum block type.

In scanning, the scanning order according to the type or direction of scanning can first be applied to sub-blocks. Additionally, a scanning order according to the direction of scanning may be applied to the quantized transform coefficients within the sub-block.

For example, as shown in FIGS. 13, 14, and 15, when the size of the target block is 8x8, the transform coefficients quantized by the first transform, second transform, and quantization of the residual signal of the target block are can be created. Thereafter, one of three scanning orders may be applied to the four 4x4 sub-blocks, and quantized transform coefficients may be scanned for each 4x4 sub-block according to the scanning order.

The encoding device 100 may generate entropy-encoded quantized transform coefficients by performing entropy encoding on the scanned quantized transform coefficients, and generate a bitstream including the entropy-encoded quantized transform coefficients. .

The decoding device 200 can extract entropy-encoded quantized transform coefficients from a bitstream and generate quantized transform coefficients by performing entropy decoding on the entropy-encoded quantized transform coefficients. Quantized transformation coefficients can be arranged in a two-dimensional block form through inverse scanning. At this time, as a reverse scanning method, at least one of (upper right) diagonal scan, vertical scan, and horizontal scan may be performed.

In the decoding device 200, dequantization may be performed on the quantized transform coefficients. Depending on whether the secondary inverse transformation is performed, the secondary inverse transformation may be performed on the result generated by performing the inverse quantization. Additionally, depending on whether the first inversion is performed, the first inversion may be performed on the result generated by performing the second inversion. A reconstructed residual signal can be generated by performing a first-order inversion on the result generated by performing a second-order inversion.

For luma components reconstructed through intra prediction or inter prediction, inverse mapping of the dynamic range may be performed before in-loop filtering.

The dynamic range can be divided into 16 equal pieces, and a mapping function for each piece can be signaled. The mapping function can be signaled at the slice level or tile group level.

A reverse mapping function for performing reverse mapping may be derived based on the mapping function.

In-loop filtering, storage of reference pictures, and motion compensation can be performed in the demapped region.

A prediction block generated through inter prediction can be converted into a mapped area by mapping using a mapping function, and the converted prediction block can be used to generate a reconstructed block. However, since intra prediction is performed in a mapped region, the prediction block generated by intra prediction can be used to generate a reconstructed block without mapping and/or demapping.

For example, if the target block is a residual block of a chroma component, the residual block can be converted to a demapped region by performing scaling on the chroma component of the mapped area.

Whether scaling is available can be signaled at the slice level or tile group level.

For example, scaling can only be applied if mapping for the luma component is available and the splitting of the luma component and the splitting of the chroma component follow the same tree structure.

Scaling may be performed based on the average of the values of samples of the luma prediction block corresponding to the chroma prediction block. At this time, if the target block uses inter prediction, the luma prediction block may mean a mapped luma prediction block.

The value required for scaling can be derived by referring to the look-up table using the index of the piece to which the average value of the samples of the luma prediction block belongs.

By performing scaling on the residual block using the finally derived value, the residual block can be converted into a demapped area. Thereafter, for the chroma component block, reconstruction, intra prediction, inter prediction, in-loop filtering, and storage of the reference picture can be performed in the demapped region.

For example, information indicating whether mapping and/or de-mapping of such luma components and chroma components is available may be signaled through a sequence parameter set.

The prediction block of the target block may be generated based on the block vector. A block vector may indicate displacement between a target block and a reference block. The reference block may be a block in the target image.

In this way, the prediction mode that generates a prediction block with reference to the target image may be called an intra block copy (IBC) mode.

IBC mode can be applied to CUs of a specified size. For example, IBC mode can be applied to MxN CU. Here, M and N may be less than or equal to 64.

IBC mode may include skip mode, merge mode, and AMVP mode. In the case of skip mode or merge mode, a merge candidate list may be constructed, and a merge index may be signaled, thereby specifying one merge candidate among the merge candidates in the merge candidate list. The block vector of the specified merge candidate can be used as the block vector of the target block.

For AMVP mode, differential block vectors can be signaled. Additionally, the prediction block vector may be derived from the left neighboring block and the top neighboring block of the target block. Additionally, an index regarding which neighboring block will be used may be signaled.

The prediction block in IBC mode may be included in the target CTU or the left CTU, and may be limited to blocks within the previously reconstructed area. For example, the value of the block vector may be limited so that the prediction block of the target block is located within a specified area. The specified area may be an area of three 64x64 blocks that are encoded and/or decoded before the 64x64 block containing the target block. By limiting the value of the block vector in this way, memory consumption and device complexity according to the implementation of the IBC mode can be reduced.

Figure 16 is a structural diagram of an encoding device according to an embodiment.

The encoding device 1600 may correspond to the encoding device 100 described above.

The encoding device 1600 includes a processing unit 1610, a memory 1630, a user interface (UI) input device 1650, a UI output device 1660, and storage that communicate with each other through a bus 1690. (1640) may be included. Additionally, the encoding device 1600 may further include a communication unit 1620 connected to the network 1699.

The processing unit 1610 may be a semiconductor device that executes processing instructions stored in a central processing unit (CPU), memory 1630, or storage 1640. The processing unit 1610 may be at least one hardware processor.

The processing unit 1610 may generate and process signals, data, or information that are input to the encoding device 1600, output from the encoding device 1600, or used inside the encoding device 1600. Inspection, comparison, and judgment related to data or information can be performed. That is, in an embodiment, generation and processing of data or information, and inspection, comparison, and judgment related to the data or information may be performed by the processing unit 1610.

The processing unit 1610 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, an entropy encoding unit 150, and an inverse quantization unit. It may include a unit 160, an inverse transform unit 170, an adder 175, a filter unit 180, and a reference picture buffer 190.

Inter prediction unit 110, intra prediction unit 120, switch 115, subtractor 125, transform unit 130, quantization unit 140, entropy encoding unit 150, inverse quantization unit 160, At least some of the inverse transform unit 170, the adder 175, the filter unit 180, and the reference picture buffer 190 may be program modules and may communicate with an external device or system. Program modules may be included in the encoding device 1600 in the form of an operating system, application program module, and other program modules.

Program modules may be physically stored on various known storage devices. Additionally, at least some of these program modules may be stored in a remote memory device capable of communicating with the encoding device 1600.

Program modules are routines, subroutines, programs, objects, components, and data that perform a function or operation according to an embodiment or implement an abstract data type according to an embodiment. It may include data structures, etc., but is not limited thereto.

Program modules may be composed of instructions or codes that are executed by at least one processor of the encoding device 1600.

The processing unit 1610 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, an entropy encoding unit 150, and an inverse quantization unit. Commands or codes of the unit 160, the inverse transform unit 170, the adder 175, the filter unit 180, and the reference picture buffer 190 can be executed.

The storage unit may represent memory 1630 and/or storage 1640. Memory 1630 and storage 1640 may be various types of volatile or non-volatile storage media. For example, the memory 1630 may include at least one of ROM 1631 and RAM 1632.

The storage unit may store data or information used for the operation of the encoding device 1600. In an embodiment, data or information held by the encoding device 1600 may be stored in the storage unit.

For example, the storage unit can store pictures, blocks, lists, motion information, inter prediction information, and bitstreams.

The encoding device 1600 may be implemented in a computer system that includes a recording medium that can be read by a computer.

The recording medium may store at least one module required for the encoding device 1600 to operate. The memory 1630 may store at least one module, and the at least one module may be configured to be executed by the processing unit 1610.

Functions related to communication of data or information of the encoding device 1600 may be performed through the communication unit 1620.

For example, the communication unit 1620 may transmit a bitstream to the decoding device 1700, which will be described later.

Figure 17 is a structural diagram of a decoding device according to an embodiment.

The decoding device 1700 may correspond to the decoding device 200 described above.

The decryption device 1700 includes a processing unit 1710, a memory 1730, a user interface (UI) input device 1750, a UI output device 1760, and storage that communicate with each other through a bus 1790. (1740). Additionally, the decryption device 1700 may further include a communication unit 1720 connected to the network 1799.

The processing unit 1710 may be a semiconductor device that executes processing instructions stored in a central processing unit (CPU), memory 1730, or storage 1740. The processing unit 1710 may be at least one hardware processor.

The processing unit 1710 may generate and process signals, data, or information that are input to the decoding device 1700, output from the decoding device 1700, or used inside the decoding device 1700. Inspection, comparison, and judgment related to data or information can be performed. That is, in an embodiment, generation and processing of data or information, and inspection, comparison, and judgment related to the data or information may be performed by the processing unit 1710.

The processing unit 1710 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, a switch 245, an adder 255, and a filter. It may include a unit 260 and a reference picture buffer 270.

Entropy decoding unit 210, inverse quantization unit 220, inverse transform unit 230, intra prediction unit 240, inter prediction unit 250, switch 245, adder 255, filter unit 260, and At least some of the reference picture buffers 270 may be program modules and may communicate with an external device or system. Program modules may be included in the decryption device 1700 in the form of an operating system, application program module, and other program modules.

Program modules may be physically stored on various known storage devices. Additionally, at least some of these program modules may be stored in a remote memory device capable of communicating with the decoding device 1700.

Program modules are routines, subroutines, programs, objects, components, and data that perform a function or operation according to an embodiment or implement an abstract data type according to an embodiment. It may include data structures, etc., but is not limited thereto.

Program modules may be composed of instructions or codes that are executed by at least one processor of the decoding device 1700.

The processing unit 1710 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, a switch 245, an adder 255, and a filter. Instructions or codes of the unit 260 and the reference picture buffer 270 may be executed.

The storage unit may represent memory 1730 and/or storage 1740. Memory 1730 and storage 1740 may be various types of volatile or non-volatile storage media. For example, the memory 1730 may include at least one of ROM 1731 and RAM 1732.

The storage unit may store data or information used for the operation of the decoding device 1700. In an embodiment, data or information held by the decoding device 1700 may be stored in the storage unit.

For example, the storage unit can store pictures, blocks, lists, motion information, inter prediction information, and bitstreams.

The decryption device 1700 may be implemented in a computer system that includes a recording medium that can be read by a computer.

The recording medium may store at least one module required for the decoding device 1700 to operate. The memory 1730 may store at least one module, and the at least one module may be configured to be executed by the processing unit 1710.

Functions related to communication of data or information of the decryption device 1700 may be performed through the communication unit 1720.

For example, the communication unit 1720 may receive a bitstream from the encoding device 1600.

Hereinafter, the processing unit may refer to the processing unit 1610 of the encoding device 1600 and/or the processing unit 1710 of the decoding device 1700. For example, for functions related to prediction, the processing unit may represent switch 115 and/or switch 245. In functions related to inter prediction, the processing unit may represent an inter prediction unit 110, a subtractor 125, and an adder 175, and may represent an inter prediction unit 250 and an adder 255. In functions related to intra prediction, the processing unit may represent an intra prediction unit 120, a subtractor 125, and an adder 175, and may represent an intra prediction unit 240 and an adder 255. In functions related to transformation, the processing unit may represent a transformation unit 130 and an inverse transformation unit 170, and may indicate an inverse transformation unit 230. In functions related to quantization, the processing unit may represent a quantization unit 140 and an inverse quantization unit 160, and may represent an inverse quantization unit 220. In functions related to entropy encoding and/or decoding, the processing unit may represent an entropy encoding unit 150 and/or an entropy decoding unit 210. In functions related to filtering, the processing unit may represent a filter unit 180 and/or a filter unit 260. For functions related to reference pictures, the processing unit may represent a reference picture buffer 190 and/or a reference picture buffer 270.

Embodiments for lightening in-loop filters and prediction deep neural networks

As deep neural network technology develops, technologies based on deep neural networks are being proposed for video compression. Video compression technologies based on deep neural networks can show superior performance than existing video compression technologies. On the other hand, video compression technologies based on deep neural networks may have the disadvantage of requiring a very large execution time compared to existing video compression technologies due to the massive computational amount of deep neural networks.

In the following embodiments, a method, device, and recording medium for efficiently lightweighting an in-loop filter and a predictive deep neural network may be provided in order to improve image encoding efficiency and shorten the time required for encoding.

In the embodiments below, a method, apparatus, and recording medium may be provided that provide learning in a lightweight in-loop filter and lightweight predictive deep neural network through knowledge distillation techniques. there is.

Through the following embodiments, a learning method in an in-loop filter and predictive deep neural network can be provided to improve the efficiency and speed of image encoding.

In embodiments, the terms “deep neural network” and “neural network” may be used interchangeably. That is to say, “in-depth” is an optional expression and may be removed or omitted in embodiments.

In embodiments, the terms “neural network” and “network” may be used interchangeably. That is to say, the term “neural network” and the term “network” may have the same meaning.

In embodiments, the terms “learning” and “training” may be used interchangeably. That is to say, the term “learning” and the term “training” may have the same meaning.

Figure 18 is a flowchart of an encoding method according to an embodiment.

The deep neural network may include one or more of an in-loop filter and a prediction deep neural network. Alternatively, the deep neural network may represent one or more of a predictive deep neural network for prediction and an in-loop filter based on the deep neural network. In embodiments, the deep neural network may be replaced with an in-loop filter and/or prediction deep neural network.

In step 1810, the processing unit 1610 of the encoding device 1600 may perform learning in a deep neural network.

At step 1820, processor 1610 may select a deep neural network.

In step 1830, the processing unit 1610 may perform encoding on deep neural network information. The processing unit 1610 may generate encoded deep neural network information by performing encoding on the deep neural network information.

In step 1840, the processor 1610 may generate a bitstream including deep neural network information or encoded deep neural network information.

The processing unit 1610 may store the bitstream in the storage unit 1630.

The communication unit 1620 may transmit a bitstream to the decoding device 1700.

Deep neural network information or encoded deep neural network information may be signaled from the encoding device 1600 to the decoding device 1700 through a bitstream.

Figure 19 is a flowchart of a decryption method according to an embodiment.

The deep neural network may include one or more of an in-loop filter and a prediction deep neural network. Alternatively, the deep neural network may represent one or more of a predictive deep neural network for prediction and an in-loop filter based on the deep neural network. In embodiments, the deep neural network may be replaced with an in-loop filter and/or prediction deep neural network.

In step 1910, the communication unit 1730 may receive a bitstream from the encoding device 1600.

In step 1920, the processing unit 1710 of the decoding device 1700 may perform learning in a deep neural network.

Steps 1910 and 1920 may be performed simultaneously. Alternatively, step 1920 may be performed before step 1910.

The storage unit 1720 can store a bitstream.

The processing unit 1710 may obtain a bitstream from the storage unit 1720 or the communication unit 1730.

The bitstream may include deep neural network information or encoded deep neural network information.

In step 1930, the processing unit 1610 may perform decoding on the encoded deep neural network information. The processing unit 1610 may generate deep neural network information by performing encoding on the encoded deep neural network information.

At step 1940, processor 1610 may select a deep neural network.

Figure 20 is a flow diagram of learning in a deep neural network according to one example.

Step 1810 may include steps 2010, 2020, and 2030 below.

Step 1920 may include steps 2010, 2020, and 2030 below.

In embodiments, learning in a deep neural network may mean learning in an in-loop filter and/or learning in a predictive deep neural network.

At step 2010, the processing unit may perform learning on a teacher network.

At step 2020, the processing unit may perform first learning on the student network. First learning may refer to the first stage of learning.

In step 2030, the processing unit may perform second learning in the student network. Secondary learning may refer to the second stage of learning.

In embodiments, the size of the input of the in-loop filter may be the same as the size of the reconstructed image or reconstructed block (before in-loop filtering is applied).

Learning in the filter described in the embodiments may be learning under specific conditions or learning for a specific target.

The specific condition may be a condition in which the value of the coding parameter is a specific value. For example, a specific condition may be that the value of the quantization parameter is a specific value. For example, the specific value could be 22, 27, 32, 37, or 42.

A specific object may be a component of an image or block. For example, the specific object may be a luma component and/or a chroma component.

A specific target may be a slice type. For example, slice types may include Intra _A , _{Intra B} , and Inter _.

Due to encoding QP fluctuations, low-latency B and Random Access (RA) can utilize intra slice filters trained using different quantization parameters. For example, the different quantization parameters may be {19, 24, 29, 34, 39}.

In embodiments, the in-loop filter may restore a reconstructed image or a reconstructed block for each component of the input image.

Figure 21 shows learning in a teacher network according to an example.

In step 2010, learning in an in-loop filter based on a deep neural network may be performed to improve the image quality of an image in which compression deterioration has occurred.

For example, learning in an in-loop filter based on a deep neural network can be performed as shown in FIG. 21.

The in-loop filter may include a filtering deep neural network. A filtering deep neural network may include multiple modules.

The input of the in-loop filter (or filtering deep neural network) may include an image in which compression deterioration has occurred and coding information of the image. An image with compression deterioration may be input to the first module among a plurality of modules. Coding information of an image may be input to each module of a plurality of modules. Coding information of an image may be input into one or more modules among a plurality of modules.

The output of the in-loop filter (or filtering deep neural network) may be an image from which compression degradation has been removed.

In embodiments, elimination of compression degradation may not mean complete elimination of compression degradation. For example, an image from which compression degradation has been removed may be an image from which compression degradation has been partially removed.

For example, an image from which compression degradation has been removed may be an image created by applying a specific process to an image from which compression degradation occurred.

For example, an image from which compression degradation has been removed may be closer to the original image than an image from which compression degradation occurred.

Additionally, an image from which compression degradation has been removed may simply represent the output of an in-loop filter, and the in-loop filter may perform processing to remove or reduce compression degradation in the image.

In embodiments, the image may refer to a target image including a target block.

In embodiments, the coding information may include the coding parameters described above. In other words, the coding information may be information related to the target image or target block described in the embodiments.

For example, the input of an in-loop filter may be an image in which compression degradation has occurred.

For example, the output of an in-loop filter may be an image from which compression degradation has been removed.

For example, learning in an in-loop filter based on a deep neural network may be performed so that the error between the output of the in-loop filter and the original image is minimized.

In embodiments, the original image may refer to an original block in the original image corresponding to the target block.

For example, learning in multiple in-loop filters may be performed, depending on various degrees of compression degradation.

For example, an in-loop filter based on a learning deep neural network may be an image reconstruction deep neural network.

For example, an image reconstruction deep neural network may include a plurality of convolutional neural network layers or fully-connected neural network layers.

For example, multiple neural network layers may be grouped into one module.

Figure 22 shows learning in an intra-prediction deep neural network and an inter-prediction deep neural network according to one embodiment.

The deep neural network may be a predictive deep neural network. A predictive deep neural network can generate a predicted image (or prediction block) similar to the original image (or original block) through prediction.

In step 2010, learning in a prediction deep neural network may be performed to generate a prediction image (or prediction block) similar to the original image (or original block) through intra prediction and/or inter prediction.

In embodiments, the term “block” may be replaced with the term “signal”. For example, the term “prediction block” may be replaced with the term “prediction signal.”

For example, learning in a predictive deep neural network can be structured as shown in FIG. 22.

The predictive deep neural network may include an intra-predictive deep neural network and/or an inter-predictive deep neural network. For example, the predictive deep neural network may be an intra-predictive deep neural network or an inter-predictive deep neural network.

The inter prediction deep neural network may include multiple modules.

The input of the intra prediction deep neural network may include coding information of the image and neighboring samples of the target block. There may be multiple peripheral samples.

The output of the intra prediction deep neural network may be an improved prediction block of the target block.

The input of the inter prediction deep neural network may include coding information of the prediction block and image of the target block. The prediction block of the target block may be input into the first module among the plurality of modules. Coding information of an image may be input to each module of a plurality of modules. Coding information of an image may be input into one or more modules among a plurality of modules.

The output of the inter prediction deep neural network may be an improved prediction block of the target block.

Alternatively, an improved prediction block of the target block may be generated using the output of the intra prediction deep neural network and the output of the inter prediction deep neural network.

For example, the input of the intra prediction deep neural network may include at least one of a neighboring sample of the target block and a prediction block of the target block. Neighboring samples may be samples adjacent to the target block. The ambient sample may be the reference sample described above in the embodiments. The surrounding sample may be a sample that has already been encoded/decoded.

For example, learning in a predictive deep neural network may be performed so that the error between the output of the predictive deep neural network and the original image is minimized.

For example, the input of an inter prediction deep neural network may be the prediction block of the current block. A prediction block may be a block in a reference picture. A prediction block can be determined through compensation using movement that minimizes rate-distortion loss. It may be a prediction block within. The reference picture may be a frame that has already been encoded/decoded.

For example, learning in multiple predictive deep neural networks may be performed, depending on various degrees of compression degradation.

For example, a predictive deep neural network may include multiple convolutional neural network layers or pre-connected neural network layers.

For example, multiple neural network layers may be grouped into one module.

Learning in teacher networks

In the above-described step 2010, the coding information of the image can be used for learning in the teacher network.

For example, as in the embodiments described above with reference to FIGS. 21 and 22 , coding information of an image used in learning in a teacher network may include coding parameters. For example, the coding information of the image includes 1) information about the block division of the image, 2) information about the encoding mode of the target block, 3) information about the encoding mode of the surrounding block, and 4) information about the quantization of the target block. and 5) information about the number of times an in-loop filter is performed on the target block.

For example, video coding information can be input into the teacher network. At this time, the coding information of the video may be input to the teacher network together with the input video or surrounding samples of the teacher network.

For example, video coding information can be input into a specific module of the teacher network. At this time, the coding information of the video may be input to a specific module along with the output of the previous module. The previous module may be a module connected to a specific module. The previous module may be a module located before a specific module.

For example, the coding information of the video may be converted to a specific location corresponding to the output of a specific module of the teacher network, and the converted coding information may be input to the specific location.

For example, video encoding information can be input to directly change model parameters within a module of the teacher network.

The learned deep neural network described in the embodiments can be utilized as a teacher network in knowledge distillation.

Additional information may be input to the filtering deep neural network and prediction deep neural network described above. Additional information may include a prediction image (or prediction block), slice quantization parameters, base quantization parameters, and slice type. Additionally, the additional information may include coding parameters.

First learning in the student network

In step 2020, a student network may be constructed with a deep neural network structure similar to the deep neural network structure of the teacher network.

In embodiments, the pre-trained teacher network may be a deep neural network of greater complexity compared to the student network on which learning will be performed. Meanwhile, the complexity of the teacher network and the complexity of the student network may be the same. The difference between the complexity of the teacher network and the complexity of the student network may not be large.

For example, the number of modules in the teacher network and the number of modules in the student network may be the same. Here, the structure of the modules of the teacher network and the structure of the modules of the student network may be the same or similar to each other.

For example, the size of the feature map of the teacher network and the size of the feature map of the student network may be the same.

For example, the size of the activity vector of the teacher network and the size of the activity vector of the student network may be the same.

Learning on the student network can be performed using knowledge distillation using a pre-trained teacher network. When learning using knowledge distillation is performed, information in the teacher network may be passed to the student network, and learning in the student network may be performed using the passed information.

Through learning on the teacher network, the teacher network can be pre-trained. Next, the student network can be trained using the Attention Transfer (AT) loss. The intermediate features of the teacher network and student network can be averaged along the channel axis to obtain one-channel attention maps. In knowledge distillation, a student network can be trained to mimic the attention map obtained from the teacher network. For effective knowledge distillation, features can be mapped to an attention map.

For example, the degree of compression degradation of the input of the teacher network and the degree of compression degradation of the input of the student network may be the same.

For example, the size of neighboring samples input to the teacher network and the size of neighboring samples input to the student network may be the same.

For example, learning in an in-loop filter based on a deep neural network may be performed until the knowledge distillation error converges through knowledge distillation.

For example, when learning using knowledge distillation is performed, no additional learning is performed on the learned teacher network, and learning may be performed only on the student network. In embodiments, learning may refer to learning within a student network.

For example, when learning using knowledge distillation is performed, learning may be performed in both the teacher network and the student network together. In embodiments, learning may refer to learning within a teacher network and learning within a student network.

For example, learning may be performed so that the values of corresponding model parameters in the deep neural network of the teacher network and the deep neural network of the student network are the same.

For example, learning may be performed so that the error between the results of the teacher network and the results of the student network is minimized.

For example, learning in a student network can be performed using intermediate feature maps or activation vectors. The intermediate feature map or activation vector may be the output of a specific module among a plurality of modules of the teacher network.

For example, a first module in a teacher network and a second module in a student network may correspond to each other. A correspondence between a first module in the teacher network and a second module in the student network may be established. Learning can be performed so that outputs from corresponding modules in the teacher network and student network have similar characteristics. Here, outputs may include intermediate feature maps or activation vectors.

For example, learning in the teacher network and the student network may be performed such that it is difficult for a separate discriminator deep neural network to distinguish between the output from the teacher network and the output from the student network. Here, the output may include a feature map, activation vector, or output image. The output image may be an image from which compression degradation has been removed. For example, feature maps or activation vectors may be extracted from modules at the same location among the overall structure of the teacher network and the overall structure of the student network. Learning may be performed using extracted feature maps or extracted learning vectors.

For example, results (eg, feature maps or activation vectors) from all modules of the teacher network and student network can be used for learning.

For example, learning may be performed such that the complexity of the student network is minimized. Here, learning can be performed such that the complexity of the student network and the knowledge distillation error are simultaneously minimized.

For example, information in the teacher network may include one or more of model parameters, intermediate feature maps, activation vectors, and filtered images. Intermediate feature maps and activation vectors may be functions of specific training images. The training image may be an input image input to the teacher network. The filtered image may be the (final) output from the teacher network.

For example, information from multiple teacher networks can be used for learning in the teacher network and student network.

For example, information from a specific teacher network among a plurality of teacher networks may be used for learning. Here, the specific teacher network may be a teacher network that generates an output with the smallest error with respect to the original image among a plurality of teacher networks.

For example, information on the teacher network may include coding information on learning videos. Information from the teacher network including coding information of the learning video can be used for learning.

Learning using knowledge distillation in embodiments may include one or more of the approaches described above.

The computational process of the attention module in relation to the attention map can be expressed as Equation 1 below.

[Formula 1]

F_out = F_in * f ( Rec , Pred , BS , QP ) + F_in

F_in may be the input of the attention module. F_out may be the output of the attention module.

Rec, Pred, BS, and QP may represent the reconstructed image, predicted image, boundary strength, and quantization parameters, respectively. The quantization parameter may be a sequence-level input quantization parameter. f may be composed of two convolutions. The activation function can be applied after the first convolution layer. The purpose of f may be to create an attention map from external information, and then perform recalibration of the feature map F_in .

Input to the neural network may include reconstructed images, predicted images, edge strengths, and quantization parameters. (Split can additionally be taken as input in intra slice models.)

It may be determined whether filtering will be applied to each slice or each block. If filtering is determined to be applied to a slice or block, conditional parameters may be additionally determined from the candidate list. The candidate list may include three candidates derived from quantization parameters.

q may represent a sequence level quantization parameter. The candidate list may include conditional parameters {q, q-5, q-10}. This selection process may be based on rate-distortion cost on the part of the encoding device 1600.

The deep neural network information may include information indicating whether to use a selection process and an index to the conditional eye parameter. The index may indicate a conditional parameter used for filtering among conditional parameters.

The granularity of the filter's determination and choice of parameters may vary based on the resolution and quantization parameters. For example, given higher resolution and larger quantization parameters, decisions and selections can be performed over larger areas.

For different temporal layers, the candidate parameter list for the temporal layer may be different. For high temporal layers, the third candidate q - 10 can be replaced by q + 5.

For all-intra configurations, parameter selection can be disabled while on/off control is still maintained.

When a neural network filter is applied to the reconstructed image, a scaling factor may be signaled for each color component in the picture header. Differences between the input image and the image filtered by the neural network may be scaled by a scaling factor before the differences are added to the input image. This scaling may be named residual scaling.

To enable combination with deblocking, the input image used in residual scaling may be the output of a deblocking filter.

Equation 2 below can represent the process of residual scaling.

[Formula 2]

R _Refine = ( R _NN - R _DB ) * w + R _DB = w * R _NN + (1 - w ) * R _DB

R _NN may be the output of a neural network filter.

R _DB may be the output of a deblocking filter.

w may be a weight.

R _Refine may be the result of residual scaling.

23 shows first learning in a student network according to one embodiment.

At step 2020, learning in a learning network using a teacher network may be performed, for learning in an in-loop filter using knowledge distillation.

For example, first learning in a student network using knowledge distillation may be performed according to the configuration shown in FIG. 23.

For example, feature maps that are the results of a module of the teacher network and a module of the student network may be extracted. Each of the extracted feature maps can be converted into an attention map. Attention maps can be created using the extracted feature maps. Here, an average can be derived for each axis of the feature map, and an attention map can be constructed using the derived average. A sum may be derived for each axis of the feature map, and an attention map may be constructed using the derived sum.

For example, the first feature map may be the result of a module of the teacher network. The second feature map may be the result of a module of the student network. When the size of the first feature map and the size of the second feature map are different from each other, an attention map can be constructed by transforming the feature map using a transformation module with parameters that can be learned.

For example, the first attention map may be an attention map obtained from the results of a module of the teacher network. The second attention map may be an attention map obtained from the results of a module of the student network. Learning may be performed so that the error between the first attention map and the second attention map is minimized.

Here, the learning loss function can be configured as Equation 3 below.

[Formula 3]

L _step1 can represent the loss function of learning.

may represent an attention map derived from the feature map output from the ith module of the teacher network.

may represent an attention map derived from the feature map output from the ith module of the student network.

24 illustrates first learning in a student network according to one embodiment.

At step 2020, learning in a learning network using a teacher network may be performed, for learning in a predictive deep neural network using knowledge distillation.

For example, first learning in a student network using knowledge distillation may be performed according to the configuration shown in FIG. 24.

For example, activation vectors that are the results of a module in the teacher network and a module in the student network can be extracted. The first activation vector may be the result of a module of the teacher network. The second activation vector may be the result of a module of the student network. A first activation vector may be extracted from a module of the teacher network. A second activation vector may be extracted from the modules of the student network. Here, a transformation module may be used to make the dimensions of the first activation vector and the dimension of the second activation vector equal to each other. Multiple transformation modules can be added to the predictive deep neural network.

For example, the transformation module may change the dimension of the second activation vector to be the same as the dimension of the first activation vector by performing transformation on the second activation vector. The transformation module may generate a second activation vector with changed dimensions by performing transformation on the second activation vector. The parameters of the transformation module can be determined through learning.

For example, learning may be performed such that the error between activation vectors from modules in the teacher network and modules in the student network is minimized. Here, the active vectors may be a first active vector and a second active vector. Alternatively, the activation vectors may be a first activation vector and a second activation vector with altered dimensions.

Here, the learning loss function can be configured as Equation 4 below.

[Formula 4]

L _step1 can represent the loss function of learning.

may represent the first activation vector output from the ith module of the teacher network.

may be a second activation vector with a changed dimension created by performing a transformation on the second activation vector output from the ith module of the student network.

25 illustrates secondary learning in a student network according to one embodiment.

In step 2030, second learning may be performed on the student network that completed the first learning. In the second learning, 1) the original image or surrounding samples, and 2) the output of the teacher network can be used.

In the second learning, fine-tuning of the student network to which the first learning is applied may be performed. Here, the output of the teacher network and the original image can be used as supervision signals for fine-tuning. Both directors can be used to measure L1 loss. Each of the two supervisors can be named as imitation loss (for the output of the teacher network) and reconstruction loss (for the original image), respectively.

The purpose of the second learning can be expressed as Equation 5 below.

[Formula 5]

L _Finetune = L _{Reconstuction} + α * L _imitation

L _Finetune can be L1 loss. L _{Reconstuction} can be a reconstruction loss. L _imitation may be imitation loss.

The effectiveness of the two loss terms can be adjusted using the hyperparameter α. α can (empirically) be set to 1.

For example, secondary learning of the student network may be performed according to the configuration shown in FIG. 25.

For example, learning in an in-loop filter and prediction deep neural network may be performed such that the error between the output of the student network and the output of the teacher network is minimized.

In embodiments, the output of the student network may be a resulting image output from the student network. The output of the teacher network may be a result image output from the teacher network.

For example, learning in an in-loop filter and prediction deep neural network can be performed so that the error between the output of the student network and the original image is minimized.

For example, learning in an in-loop filter and prediction deep neural network may be performed such that the first error and the second error are simultaneously minimized. The first error may be an error between the output of the student network and the output of the teacher network. The second error may be an error between the output of the student network and the original image. Here, the proportions of the first error and the second error in learning can be adjusted by a Lagrange multiplier. For example, the value of the Lagrange multiplier may be set to 1.

For example, learning in an in-loop filter and prediction deep neural network may be performed until the first error and the second error converge to each other. Here, the learning loss function can be configured as Equation 6 below.

[Formula 6]

L _step2 can represent the loss function of learning.

may represent the first error.

may represent the second error.

For example, a first loss function of a first learning in the student network and a second loss function of a second learning in the student network can be combined into one third loss sum. Learning in the in-loop filter and prediction deep neural network may be performed using a third loss function.

Choosing a Deep Neural Network

At steps 1820 and 1940, a deep neural network may be selected. Selecting a deep neural network may mean performing filtering on the target image using the selected deep neural network. Here, the target image may be a reconstructed image. The target image may mean a target block or a reconstructed block.

In steps 1820 and 1940, filtering may be performed on the target image using a filter based on a learned deep neural network. The deep neural network filter may be a learned deep neural network described in the embodiments. Alternatively, the deep neural network filter may be a filter that uses the learned deep neural network described in the embodiments.

In embodiments, filter may refer to a deep neural network filter.

When filtering is performed on a target image, one filter among N filters may be selected. Here, the N filters may be N deep neural network filters. N deep neural network filters can each perform different degrees of image reconstruction. That is, N deep neural network filters may differ from each other in the degree of image reconstruction.

For example, the N deep neural network filters may be student networks learned using knowledge distillation.

For example, it may be determined whether a filter is applied to a specific target. For example, a specific target may be a unit containing a target block.

In embodiments, the specific object is a (reconstructed) image, (reconstructed) picture, (reconstructed) slice, (reconstructed) Coding Tree Unit (CTU) or (reconstructed) It can be a block.

Here, the size of the (rebuilt) block may be HxW. N, H and W can be positive integers.

For example, when filtering is performed on a specific target, one filter among N filters may be selected.

For example, a set of filter candidates may be configured according to the degree of compression and compression method for a specific object (or image). The filter candidate set may be some of the N filters. In selecting the filter described in the embodiments, one filter may be selected from among the filters in the filter candidate set. The filter candidate set may be M deep neural network filters. M may be an integer less than or equal to N. The M deep neural network filters in the filter candidate set can each perform different degrees of image reconstruction.

For example, the filter with the smallest error among N filters or filters in the filter candidate set may be selected. Here, the error may be the error between the resulting image of the filter and the original image.

For example, a filter may be selected using the results of filtering on a reference unit. For example, a reference unit may include neighboring blocks, CTUs, slices, and reference images adjacent to the target image.

In embodiments, the filter's weight α may be used for the filter.

For example, one weight may be used or different weights may be used depending on the characteristics of the target image. The target image may be an input image that is subject to compression.

For example, deep neural network information may include weight information. Weight information may represent weight α. Weight information may be signaled through a bitstream.

For example, the number of weights that change can be determined based on the characteristics of the target image. The number of weights can be determined in terms of rate-distortion in the image compression process.

For example, the output of a filter can be adjusted by using weights to minimize the error between the filter's output and the original image. Equation 7 below can represent an example of adjusting the output of a filter using weights.

[Formula 7]

I _input may be the original video. I _filtered may be the output of a filter. I _output may be a reconstructed image. I _output may be the reconstructed image with an (in-loop) filter applied. α may be a weight.

For example, a weighted sum of the output of a deep neural network filter and the results of a conventional in-loop filter can be used as the reconstructed image. At this time, one or more results of the existing in-loop filter may be used for the weighted sum. The result of the in-loop filter may mean a reconstructed image to which the in-loop filter has been applied. Equation 8 below can represent a weighted sum using the output of a deep neural network filter.

[Formula 8]

I _CNN may be the output of a deep neural network filter. I _LF may be the result of an in-loop filter. I _output may be a reconstructed image. I _output may be the reconstructed image with an (in-loop) filter applied. α may be a weight.

For example, the weight used for the output of a deep neural network filter for a specific target can be adjusted. For example, a specific target may be a unit containing a target block.

In embodiments, the specific object may be a (reconstructed) image, (reconstructed) picture, (reconstructed) slice, (reconstructed) (CTU) or (reconstructed) block.

For example, for a particular subject, a weighted sum of the output of a deep neural network filter and the results of a conventional in-loop filter may be used.

For example, for a particular subject, the weights for a weighted sum of the output of a deep neural network filter and the results of a conventional in-loop filter may be adjusted.

For example, for a specific target, among the output of a deep neural network filter and the result of a conventional in-loop filter, the one that is more suitable in terms of rate-distortion can be selected as the reconstructed image (with the in-loop filter applied). there is.

For example, the results of an in-loop filter may be multiple. Among the plurality of results of the in-loop filter, the one that is more suitable in terms of rate-distortion may be selected as the reconstructed image (to which the in-loop filter is applied).

For example, one or more of the plurality of results of an existing in-loop filter may be used as candidates for the selection described above.

In embodiments, existing in-loop filters include a deblocking filter, Sample Adaptive Offset (SAO), Adaptive Loop Filter (ALF), and Cross Component Adaptive Loop Filter ( It may include Cross Component Adaptive Loop Filter (CCALF), etc. However, the in-loop filters of embodiments are not limited to deblocking filters, SAO, ALF and CCALF, etc. The existing in-loop filter may further include the in-loop filter of the embodiments and other in-loop filters.

For example, deep neural network filters can be applied before traditional in-loop filters. The deep neural network filter may be applied to the specific target before the existing in-loop filter is applied to the specific target. Deep neural network filters can be applied later than traditional in-loop filters. The deep neural network filter may be applied to the specific target after the existing in-loop filter is applied to the specific target. Deep neural network filters can be applied in between existing in-loop filters. The deep neural network filter may be applied to the specific object after the existing first in-loop filter is applied to the specific object, and after the deep neural network filter is applied to the specific object, the existing second in-loop filter may be applied to the specific object. Can be applied to specific targets.

For example, coding information of an image may be used in selecting a deep neural network filter.

In embodiments, one or more of the methods for selecting a deep neural network filter described above may be combined.

Filtering of reconstructed images using determined deep neural network filters

The deep neural network filter to be applied to the reconstructed image can be determined through selection of the deep neural network filter described above. In embodiments, “selected filter” may mean “determined filter.”

Hereinafter, a deep neural network may refer to a deep neural network filter selected/determined to be applied to the reconstructed image.

Hereinafter, the reconstructed image may be replaced by the specific unit described above. For example, a “reconstructed image” can be considered a (reconstructed) image, (reconstructed) slice, (reconstructed) CTU or (reconstructed) block.

Filtering may be performed on the reconstructed image using the determined deep neural network filter.

For example, the input of a deep neural network filter may be the reconstructed image generated within the loop. The input of a deep neural network filter may include the reconstructed image generated within the loop.

For example, the input of the deep neural network filter may be the luma component and/or the chroma component of the reconstructed image generated within the loop. The input of the deep neural network filter may include luma components and/or chroma components of the reconstructed image generated within the loop.

In embodiments, the size of the luma component may be larger than the size of the chroma component. Up-sampling may be performed on the chroma component to match the size of the chroma component to the size of the luma component. When the chroma component is input into a deep neural network filter or prediction deep neural network, up-sampling on the chroma component may be performed. Up-sampling on chroma components may be performed as pre-processing for filtering by a deep neural network filter or prediction by a predictive deep neural network. Once the chroma component is output from a deep neural network filter or prediction deep neural network, down-sampling on the chroma component may be performed. Down-sampling on chroma components may be performed as pre-processing for filtering by a deep neural network filter or prediction by a predictive deep neural network.

In embodiments, the chroma component may be a Cb component and/or a Cr component.

For example, when filtering is performed on the chroma component of the reconstructed image by a deep neural network filter, the luma component of the reconstructed image along with the chroma component of the reconstructed image may be input to the deep neural network filter.

For example, when filtering is performed on the chroma component of an image reconstructed by a deep neural network filter, the luma component of the reconstructed image may be included in the input of the deep neural network filter.

In embodiments, the luma component may be one of a predicted luma component (i.e., a luma component of the predicted image), a reconstructed luma component (i.e., a luma component of the reconstructed image), and a filtered reconstructed luma component. there is. The filtered reconstructed luma component may be a luma component (of the reconstructed image) to which in-loop filtering has been applied by a deep neural network filter or a conventional in-loop filter.

For example, the input of a deep neural network filter may be the filtered reconstructed image generated within the loop. A filtered reconstructed image can be generated by applying (in-loop) filtering to the reconstructed image. The input of the deep neural network filter may include the filtered and reconstructed image generated within the loop.

For example, deep neural network information may include index information indicating a filter selected by the encoding device 1600. Index information may be signaled through a bitstream. The processing unit 1710 of the decoding device 1700 can identify a filter selected from a plurality of filters using index information and can perform filtering on the reconstructed image using the selected filter.

For example, the processing unit 1710 of the decoding device 1700 may estimate or restore index information using other information such as deep neural network information and prediction information. The processing unit 1710 of the decoding device 1700 can identify the selected filter indicated by the index information among the plurality of filters and perform filtering on the reconstructed image using the selected filter. The processing unit 1710 of the decoding device 1700 may identify a selected filter indicated by index information among a plurality of filters, and may perform in-loop filtering using a deep neural network using the selected filter.

Prediction using a trained prediction deep neural network

Intra prediction and/or inter prediction may be performed using the learned prediction deep neural network described above. When intra prediction and/or inter prediction is performed, one filter out of N filters may be selected.

N can be 1, 2 or 3. Alternatively, N may be a positive integer.

The N filters may be N predictive deep neural networks. N predictive deep neural networks can use neighboring samples of the target block as input to different degrees. That is, N predictive deep neural networks may differ from each other in terms of the extent to which they use neighboring samples as input. For example, N predictive deep neural networks can use different numbers of neighboring samples as input.

For example, N predictive deep neural networks can be student networks that have been trained using knowledge distillation.

For example, whether to apply a filter may be determined for a specific target.

In embodiments, specific objects may include images, pictures, slices, CTUs, and blocks.

Here, the size of the block may be HxW. H and W can be positive integers.

For example, when intra prediction and/or inter prediction is performed for a specific target, one of N prediction deep neural networks may be selected.

For example, a set of predictive deep neural network candidates may be constructed depending on the degree of compression and compression method for a specific object (or image). The predictive deep neural network candidate set may be some of the N predictive deep neural networks. In selecting the predictive deep neural network described in the embodiments, one predictive deep neural network may be selected from among the predictive deep neural networks in the predictive deep neural network candidate set. The set of predictive deep neural network candidates may be M predictive deep neural networks. M may be an integer less than or equal to N. The M predictive deep neural networks in the predictive deep neural network candidate set can each perform different predictions.

For example, a prediction deep neural network may be selected based on the size of a block among N prediction deep neural networks or prediction deep neural networks in a set of prediction deep neural network candidates.

For example, the prediction deep neural network with the smallest error among the N prediction deep neural networks or the prediction deep neural networks in the prediction deep neural network candidate set may be selected. Here, the error may be the error between the original image and the result of prediction by the prediction deep neural network (say, an improved prediction block).

For example, a predictive deep neural network may be selected using the results of prediction or reconstruction for reference units. For example, a reference unit may include neighboring blocks, CTUs, slices, and reference images adjacent to the target image.

For example, the weights for the prediction deep neural network may change based on the characteristics of the target image. In other words, weights for the prediction deep neural network can be determined based on the characteristics of the target image. The target image may be an input image that is subject to compression.

For example, deep neural network information may include weight information. Weight information may represent weights for a predictive deep neural network. Weight information may be signaled through a bitstream.

For example, the number of weights that change can be determined based on the characteristics of the target image. The number of weights can be determined in terms of rate-distortion in the image compression process.

For example, a predictive deep neural network may be selected using the coding information of the image.

In embodiments, one or more of the methods for selecting a predictive deep neural network described above may be combined.

Intra prediction and/or inter prediction may be performed using the determined prediction deep neural network.

For example, when intra prediction is performed, the input of the prediction deep neural network may be neighboring samples of the target block. When intra prediction is performed, the input of the prediction deep neural network may include neighboring samples of the target block.

For example, when inter prediction is performed, the input of the prediction deep neural network may be at least one of an image to which inter prediction has been applied and neighboring samples of the target block. The input of the prediction deep neural network may include at least one of an image to which inter prediction has been applied and neighboring samples of the target block.

For example, when intra prediction and/or inter prediction is performed, the input of the prediction deep neural network may be the luma component and/or chroma component of the image. The input of the predictive deep neural network may include luma and/or chroma components of the image.

For example, in intra prediction and/or inter prediction, when prediction of the chroma component of an image is performed by a prediction deep neural network, the luma component of the image together with the chroma component of the image may be input to the prediction deep neural network. .Can be input into a deep neural network filter.

For example, in intra prediction and/or inter prediction, when prediction of the chroma component of an image is performed by a prediction deep neural network, the luma component of the image may be included in the input of the prediction deep neural network.

For example, the deep neural network information may include index information indicating the predictive deep neural network selected by the encoding device 1600. Index information may be signaled through a bitstream. The processing unit 1710 of the decoding device 1700 may use the index information to identify a prediction deep neural network selected from a plurality of prediction deep neural networks, and use the selected prediction deep neural network to provide intra information about the target block (or image). Prediction and/or inter prediction may be performed.

For example, the processing unit 1710 of the decoding device 1700 may estimate or restore index information using other information such as deep neural network information and prediction information. A predictive deep neural network used to perform predictions using the index information may be specified. The processing unit 1710 of the decoding device 1700 may identify the selected prediction deep neural network indicated by the index information among a plurality of prediction deep neural networks, and perform intra prediction for the target block (or image) using the selected prediction deep neural network. And/or inter prediction may be performed.

For example, in intra prediction and/or inter prediction, when prediction of the chroma component of the target block (or image) is performed by a prediction deep neural network, the luma component of the target block (or image) is predicted deep. It can be included in the input of a neural network.

In embodiments, the luma component may be a predicted luma component (i.e., a luma component of a prediction block or a predicted image), a reconstructed luma component (i.e., a luma component of a reconstructed block or a reconstructed image), and a filtered reconstruction. It may be one of the built luma components. The filtered reconstructed luma component may be a luma component (of the reconstructed block or reconstructed image) to which in-loop filtering has been applied by a deep neural network filter or a conventional in-loop filter.

Encoding/decoding of deep neural network information

In step 1830, encoding of the deep neural network information may be performed, and in step 1930, decoding of the (encoded) deep neural network information may be performed.

In embodiments, encoding may refer to entropy encoding. Decryption may mean entropy decryption.

Deep neural network information may be encoded and/or decoded within specific information within the bitstream. For example, specific information may include a sequence parameter set, a picture parameter set, and a slice header. Deep neural network information may be encoded and/or decoded between CTU syntax elements in the bitstream.

Deep neural network information may be information used to construct a deep neural network. Deep neural network information may be information representing a deep neural network. Deep neural network information may be information representing characteristics of a deep neural network.

The deep neural network information may include information for performing filtering using a filter and information for performing prediction using a predictive deep neural network.

Deep neural network information may include in-loop filter information and predictive deep neural network information.

The in-loop filter information includes 1) filtering performance information for the luma component, 2) filtering performance information for the chroma component, 3) filter selection information for the luma component, 4) filter selection information for the chroma component, and 5) specific target. Contains one or more of the following: filtering performance information for, 6) filter selection information for a specific target, 7) deep neural network filter weight change information, 8) deep neural network filter output adjustment information, and 9) deep neural network filter weighted sum information. can do.

1) Filtering performance information on the luma component may indicate whether filtering is performed on the luma component.

2) Filtering performance information for the chroma component may indicate whether filtering is performed on the chroma component.

3) Filter selection information for the luma component may indicate the filter selected for the luma component.

4) Filter selection information for the chroma component may indicate the filter selected for the chroma component.

5) Filtering performance information for a specific target may indicate whether filtering is performed for a specific target.

In embodiments, specific objects may include images, pictures, slices, CTUs, and blocks.

6) Filter selection information for a specific target may indicate a filter selected for a specific target.

7) The deep neural network filter weight change information may be information related to the weight of the deep neural network filter. The deep neural network filter weight change information may include weight information for the deep neural network filter and may include information indicating the number of weights.

8) The deep neural network filter output adjustment information may include information used to adjust the output of the deep neural network filter. The deep neural network filter output adjustment information may include information indicating whether to adjust the output of the deep neural network filter. Deep neural network filter output adjustment information may include weights.

9) The deep neural network filter weighted sum information may include information used to derive a weighted sum of the output of the deep neural network filter and the results of the existing in-loop filter. The deep neural network filter weighted sum information may include weights of the weighted sum. The deep neural network filter weighted sum information may include information indicating the result used for the weighted sum among the results of the existing in-loop filter.

Processing of at least one of filtering performance information for luma components, filter selection information for luma components, filtering performance information for chroma components, and filter selection information for chroma components includes filtering performance information for a specific target and filter selection information for a specific target. It may be performed based on filter selection information.

For example, when the value of the filtering performance information for a specific target is the first value, filtering may be performed for the specific target. If the value of the filtering performance information for a specific target is the second value, filtering may not be performed for the specific target. At this time, filtering performance information for a specific target may be signaled for each specific target.

For example, when the value of the filtering performance information for a specific target is the first value, filter selection information for the specific target may be additionally signaled. Filtering on a specific target may be performed using the selected filter indicated by the filter selection information.

The predictive deep neural network information includes 1) predictive deep neural network performance information for the luma component, 2) predictive deep neural network performance information for the chroma component, 3) predictive deep neural network selection information for the luma component, and 4) predictive deep neural network information for the chroma component. It may include one or more of selection information, 5) prediction deep neural network performance information for a specific target, 6) prediction deep neural network selection information for a specific target, and 7) prediction deep neural network weight change information.

1) The predictive deep neural network performance information for the luma component may indicate whether prediction using the predictive deep neural network is performed for the luma component.

2) The predictive deep neural network performance information for the chroma component may indicate whether prediction using the predictive deep neural network is performed for the chroma component.

3) The prediction deep neural network selection information for the luma component may indicate the prediction deep neural network selected for the luma component.

4) The prediction deep neural network selection information for the chroma component may indicate the prediction deep neural network selected for the chroma component.

5) Predictive deep neural network performance information for a specific target may indicate whether prediction using a predictive deep neural network is performed for a specific target.

In embodiments, specific objects may include CTUs and blocks.

6) Predictive deep neural network selection information for a specific target may indicate a predictive deep neural network selected for a specific target.

7) The prediction deep neural network weight change information may be information related to the weight of the prediction deep neural network. The prediction deep neural network weight change information may include weight information for the prediction deep neural network and may include information indicating the number of weights.

Processing of at least one of predictive deep neural network performance information for luma components, predictive deep neural network selection information for luma components, predictive deep neural network performance information for chroma components, and predictive deep neural network selection information for chroma components is performed on a specific target. It may be performed based on prediction deep neural network performance information and prediction deep neural network selection information for a specific target.

For example, when the value of the prediction deep neural network performance information for a specific target is the first value, prediction using the predictive deep neural network may be performed for the specific target. If the value of the prediction deep neural network performance information for a specific target is the second value, prediction using the predictive deep neural network may not be performed for the specific target. At this time, predictive deep neural network performance information for a specific target may be signaled.

For example, when the value of the predictive deep neural network performance information for a specific target is the first value, the predictive deep neural network selection information for the specific target may be additionally signaled. Prediction deep neural network selection A prediction for a specific target can be made using a selected predictive deep neural network indicated by information.

For example, deep neural network information may include scaling information. A scaling operation may be applied to the output of a filter or prediction deep neural network to refine the output of the filter or prediction deep neural network. Fixed weights can be used to blend the output of a deep neural network filter and the output of a traditional in-loop filter. For example, the weight for a deep neural network filter could be 1, 0.75, or 0.5.

The above embodiments may be performed in the encoding device 1600 and the decoding device 1700 using the same method and/or a corresponding method. Additionally, a combination of one or more of the above embodiments may be used in encoding and/or decoding an image.

The order in which the above embodiments are applied may be different in the encoding device 1600 and the decoding device 1700. Alternatively, the order in which the above embodiments are applied may be (at least partially) the same in the encoding device 1600 and the decoding device 1700.

The above embodiments can be performed for each of the luma signal and the chroma signal. The above embodiments can be performed in the same way for luma signals and chroma signals.

The shape of the block to which the above embodiments are applied may have a square shape or a non-square shape.

Whether to apply and/or perform at least one of the above embodiments may be determined based on conditions regarding the size of the block. That is, at least one of the above embodiments can be applied and/or performed when the conditions for the size of the block are met. Conditions may include minimum block size and maximum block size. The block may be one of the blocks described above in the embodiments and the units described above in the embodiments. The block to which the minimum block size is applied and the block to which the maximum block size is applied may be different.

For example, when the size of a block is greater than or equal to the minimum size and/or when the size of the block is less than or equal to the maximum size, the above-described embodiments may be applied and/or performed. If the size of the block is larger than the minimum size and/or if the size of the block is less than or equal to the maximum size, the above-described embodiments may be applied and/or performed.

For example, the above-described embodiment can be applied only when the block size is a predefined block size. Predefined block sizes can be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. _The predefined block size may be (2* _{SIZE SIZE} SIZE _Y may be one of integers greater than or equal to 1.

For example, the above-described embodiment can be applied only when the block size is greater than or equal to the minimum block size. The above-described embodiment can be applied only when the size of the block is larger than the minimum block size. Block minimum sizes can be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. Alternatively, the minimum block size may be (2*SIZE _{MIN_X} )x(2*SIZE _{MIN_Y} ). SIZE _{MIN_X} can be one of integers greater than 1. SIZE _{MIN_Y} can be one of integers greater than 1.

For example, the above-described embodiment can be applied only when the block size is less than or equal to the maximum block size. The above-described embodiment can be applied only when the block size is smaller than the maximum block size. The maximum block size can be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. Alternatively, the maximum block size may be (2*SIZE _{MAX_X} )x(2*SIZE _{MAX_Y} ). SIZE _{MAX_X} can be one of integers greater than 1. SIZE _{MAX_Y} can be one of integers greater than 1.

For example, the above-described embodiment can be applied only when the block size is greater than or equal to the minimum block size and less than or equal to the maximum block size. The above-described embodiment can be applied only when the block size is greater than the minimum block size and less than or equal to the maximum block size. The above-described embodiment can be applied only when the block size is greater than or equal to the minimum block size and smaller than the maximum block size. The above-described embodiment can be applied only when the block size is larger than the minimum block size and smaller than the maximum block size.

In the above-described embodiments, the size of the block may mean the horizontal size of the block or the vertical size of the block. The size of the block may refer to both the horizontal size of the block and the vertical size of the block. Additionally, the size of the block may mean the area of the block. Each of the area, minimum block size, and maximum block size can be one of integers greater than or equal to 1. Additionally, the size of the block may mean the result (or value) of a known formula using the horizontal and vertical sizes of the block or the result (or value) of a formula in an embodiment.

Additionally, in the above embodiments, the first embodiment may be applied to the first size, and the second embodiment may be applied to the second size.

The above embodiments can be applied according to the temporal layer. A separate identifier may be signaled to identify the temporal layer to which the above embodiments can be applied, and the above embodiments may be applied to the temporal layer specified by the identifier. The identifier here may be defined as the lowest layer and/or highest layer to which the above embodiment is applicable, or may be defined to indicate a specific layer to which the above embodiment is applicable. Additionally, a fixed temporal hierarchy to which the above embodiments are applied may be defined.

For example, the above embodiments can be applied only when the temporal layer of the target image is the lowest layer. For example, the above embodiments can be applied only when the temporal layer identifier of the target image is 1 or more. For example, the above embodiments can be applied only when the temporal layer of the target image is the highest layer.

A slice type or tile group type to which the above embodiments are applied may be defined, and the above embodiments may be applied depending on the corresponding slice type or tile group type.

In the above-described embodiments, in applying a specified process to a specified object, specified conditions may be required, and when it is described that the specified processing is processed under a specified decision, the specified coding parameters If it has been described that it is determined whether a specified condition is satisfied or that a specified decision is made based on a specified coding parameter, the specified coding parameter may be interpreted as being replaceable with another coding parameter. That is, coding parameters affecting specified conditions or specified decisions may be considered merely exemplary, and combinations of one or more other coding parameters in addition to the specified coding parameters will be understood to play the role of the specified coding parameters. You can.

In the above-described embodiments, the methods are described based on flowcharts as a series of steps or units, but the present invention is not limited to the order of steps, and some steps may occur in a different order or simultaneously with other steps as described above. You can. Additionally, a person of ordinary skill in the art will recognize that the steps shown in the flowchart are not exclusive and that other steps may be included or one or more steps in the flowchart may be deleted without affecting the scope of the present invention. You will understand.

The above-described embodiments include examples of various aspects. Although not all possible combinations for representing the various aspects can be described, those skilled in the art will recognize that other combinations are possible in addition to those explicitly described. Accordingly, the present invention is intended to include all other substitutions, modifications and changes falling within the scope of the following claims.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in the computer software field.

A computer-readable recording medium may contain information used in embodiments according to the present invention. For example, a computer-readable recording medium may include a bitstream, and the bitstream may include information described in embodiments according to the present invention.

A bitstream may contain computer-executable code and/or programs. Computer-executable code and/or program may include information described in the embodiments and may include syntax elements described in the embodiments. That is, the information and syntax elements described in the example may be considered computer-executable code within the bitstream, and may be considered at least part of the computer-executable code and/or program represented by the bitstream. Recording media may include non-transitory computer-readable medium.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and perform program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The above hardware devices may be configured to operate as one or more software modules to perform processing according to the invention and vice versa.

In the above, the present invention has been described with specific details such as specific components and limited embodiments and drawings, but this is only provided to facilitate a more general understanding of the present invention, and the present invention is not limited to the above embodiments. No, those skilled in the art can make various modifications and changes based on this description.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments, and the scope of the patent claims described later as well as all modifications equivalent to or equivalent to the scope of the claims are within the scope of the spirit of the present invention. It will be said that it belongs.

Claims (20)

1. selecting a deep neural network;

   Generating encoded deep neural network information by performing encoding on deep neural network information used to construct the deep neural network; and

   Generating a bitstream containing the encoded deep neural network information

   A video encoding method including.
2. According to paragraph 1,

   Step of performing learning in the deep neural network

   A video encoding method further comprising:
3. According to paragraph 2,

   The step of performing learning in the deep neural network is,

   performing learning in the teacher network;

   performing first learning on the student network; and

   Performing second learning in the student network

   A video encoding method including.
4. According to paragraph 3,

   A video encoding method in which learning in the student network is performed using knowledge distillation using the teacher network.
5. According to paragraph 1,

   An image encoding method in which learning is performed in an in-loop filter based on the deep neural network.
6. According to clause 5,

   The input to the in-loop filter is an image in which compression deterioration has occurred,

   The output of the in-loop filter is an image from which compression degradation has been removed,

   An image encoding method in which learning is performed in the in-loop filter so that the error between the output of the in-loop filter and the original image is minimized.
7. According to paragraph 1,

   The deep neural network is an image encoding method that is a predictive deep neural network that generates a predicted image similar to the original image through prediction.
8. Receiving a bitstream containing encoded deep neural network information;

   Generating deep neural network information used to construct a deep neural network by decoding the encoded deep neural network information; and

   Step of selecting the deep neural network

   A video decoding method including.
9. According to clause 8,

   Step of performing learning in the deep neural network

   A video decoding method further comprising:
10. According to clause 9,

    The step of performing learning in the deep neural network is,

    performing learning in the teacher network;

    performing first learning on the student network; and

    Performing second learning in the student network

    A video decoding method including.
11. According to clause 10,

    A video decoding method in which learning in the student network is performed using knowledge distillation using the teacher network.
12. According to clause 8,

    An image decoding method in which learning is performed in an in-loop filter based on the deep neural network.
13. According to clause 12,

    The input of the in-loop filter is an image in which compression deterioration has occurred,

    The output of the in-loop filter is an image from which compression degradation has been removed,

    An image decoding method in which learning is performed in the in-loop filter so that the error between the output of the in-loop filter and the original image is minimized.
14. According to clause 8,

    The deep neural network is an image decoding method that is a predictive deep neural network that generates a predicted image similar to the original image through prediction.
15. In the computer-readable recording medium storing a bitstream for video decoding, the bitstream includes:

    Encoded deep neural network information

    Including,

    By performing decoding on the encoded deep neural network information, deep neural network information used to construct a deep neural network is generated,

    A computer-readable recording medium from which the deep neural network is selected.
16. According to clause 16,

    A computer-readable recording medium on which learning in the deep neural network is performed.
17. According to clause 16,

    Learning in the teacher network is performed,

    First learning in the student network is performed,

    A computer-readable recording medium on which second learning is performed in the student network.
18. According to clause 17,

    A computer-readable recording medium wherein learning in the student network is performed using knowledge distillation using the teacher network.
19. According to clause 15,

    A computer-readable recording medium on which learning is performed in an in-loop filter based on the deep neural network.
20. According to clause 12,

    The input to the in-loop filter is an image in which compression deterioration has occurred,

    The output of the in-loop filter is an image from which compression degradation has been removed,

    A computer-readable recording medium in which learning is performed in the in-loop filter so that an error between the output of the in-loop filter and the original image is minimized.

Images