KR20240025428A - Image decoding apparatus and image encoding apparatus for adaptive quantization and inverse-quantization, and methods thereby - Google Patents

Abstract

A method of decoding an image according to an embodiment includes obtaining second feature data for first feature data obtained through neural network-based encoding of a current image from a bitstream; Obtaining quantization data and probability data by applying the second feature data to a neural network; modifying probability data based on sample values of quantization data; Obtaining quantized first feature data by applying entropy decoding based on modified probability data to bits included in the bitstream; Obtaining dequantized first feature data by dequantizing the quantized first feature data according to sample values of the quantization data; And it may include restoring the current image through neural network-based decoding of the dequantized first feature data.

Description

Image decoding device, image encoding device and method for adaptive quantization and inverse quantization {IMAGE DECODING APPARATUS AND IMAGE ENCODING APPARATUS FOR ADAPTIVE QUANTIZATION AND INVERSE-QUANTIZATION, AND METHODS THEREBY}

This disclosure relates to encoding and decoding of images. More specifically, the present disclosure relates to technology for encoding and decoding images using AI (Artificial Intelligence), for example, a neural network.

In codecs such as H.264 AVC (Advanced Video Coding) and HEVC (High Efficiency Video Coding), the video is divided into blocks, and each block is predicted, encoded, and processed through inter prediction or intra prediction. Predictive decoding is possible.

Intra prediction is a method of compressing images by removing spatial redundancy within the images, and inter prediction is a method of compressing images by removing temporal redundancy between images.

As a representative example of inter prediction, motion estimation coding exists. Motion estimation coding predicts blocks of the current image using a reference image. A reference block most similar to the current block can be searched within a predetermined range using a predetermined evaluation function. The current block is predicted based on the reference block, and the prediction block generated as a result of prediction is subtracted from the current block to generate and encode a residual block.

To derive a motion vector pointing to a reference block in a reference image, motion vectors of previously encoded blocks can be used as a motion vector predictor of the current block. The differential motion vector, which is the difference between the motion vector of the current block and the motion vector predictor, is signaled to the decoder through a predetermined method.

Recently, technologies for encoding/decoding images using AI (Artificial Intelligence) have been proposed, and a method for effectively encoding/decoding images using AI, for example, neural networks, is required.

A method of decoding an image according to an embodiment may include obtaining second feature data for first feature data obtained through neural network-based encoding of a current image from a bitstream.

In one embodiment, the image decoding method may include obtaining quantization data and probability data by applying second feature data to a neural network.

In one embodiment, a method of decoding an image may include modifying probability data based on sample values of quantized data.

In one embodiment, a method of decoding an image may include obtaining quantized first feature data by applying entropy decoding based on modified probability data to bits included in a bitstream.

In one embodiment, a method of decoding an image may include dequantizing quantized first feature data according to sample values of quantized data to obtain dequantized first feature data.

In one embodiment, the image decoding method may include restoring the current image through neural network-based decoding of dequantized first feature data.

An image encoding method according to an embodiment includes the step of applying first feature data obtained through neural network-based encoding of a current image to a first neural network to obtain second feature data for the first feature data. can do.

In one embodiment, a method of encoding an image may include obtaining quantization data and probability data by applying second feature data to a second neural network.

In one embodiment, a method of encoding an image may include modifying probability data based on sample values of quantized data.

In one embodiment, a method of encoding an image may include quantizing first feature data according to sample values of quantized data to obtain quantized first feature data.

In one embodiment, the image encoding method includes generating a bitstream including bits corresponding to the quantized first feature data by applying entropy coding based on modified probability data to the quantized first feature data. It can be included.

In one embodiment, the bitstream may include bits corresponding to second characteristic data.

An image decoding device according to an embodiment may include an acquisition unit that acquires second feature data for first feature data obtained through neural network-based encoding of a current image from a bitstream.

In one embodiment, the image decoding device may include a prediction decoder that restores the current image through neural network-based decoding of the dequantized first feature data.

In one embodiment, the acquisition unit may include an AI control unit that acquires quantization data and probability data by applying the second feature data to a neural network, and modifies the probability data based on sample values of the quantization data.

In one embodiment, the acquisition unit may include an entropy decoding unit that obtains quantized first feature data by applying entropy decoding based on modified probability data to bits included in the bitstream.

In one embodiment, the acquisition unit may include an inverse quantization unit that inversely quantizes the quantized first feature data according to sample values of the quantization data to obtain the inverse quantized first feature data.

An image encoding device according to an embodiment may include a prediction encoding unit that obtains first feature data through neural network-based encoding of the current image.

In one embodiment, a video encoding device may include a generator that generates a bitstream corresponding to the current video.

In one embodiment, the generator applies the first feature data to the first neural network to obtain second feature data for the first feature data, and applies the second feature data to the second neural network to obtain quantization data and probability data. It may include an AI control unit that acquires and modifies probability data based on sample values of quantization data.

In one embodiment, the generator may include a quantization unit that quantizes the first feature data according to sample values of the quantization data to obtain quantized first feature data.

In one embodiment, the generator may include an entropy encoding unit that generates a bitstream including bits corresponding to the quantized first feature data by applying entropy coding based on modified probability data to the quantized first feature data. You can.

In one embodiment, the bitstream may include bits corresponding to second characteristic data.

FIG. 1 is a diagram illustrating a process of encoding and decoding a current image based on intra prediction according to an embodiment.
Figure 2 is a diagram for explaining the encoding and decoding process of a current video based on inter prediction according to an embodiment.
FIG. 3 is a diagram illustrating a method of obtaining probability data used for entropy encoding and entropy decoding according to an embodiment.
FIG. 4 is a diagram illustrating a method of obtaining quantized data used for quantization and dequantization, and modified probability data used for entropy encoding and entropy decoding, according to an embodiment.
FIG. 5 is a diagram illustrating the configuration of a video decoding device according to an embodiment.
Figure 6 is a diagram illustrating the configuration of an acquisition unit according to an embodiment.
Figure 7 is a diagram for explaining the operation of the AI control unit according to one embodiment.
Figure 8 is a diagram showing the structure of a neural network according to one embodiment.
Figure 9 is a diagram for explaining a method of modifying probability data according to an embodiment.
Figure 10 is a diagram showing probability data, modified probability data, and probability model according to one embodiment.
Figure 11 is a diagram for explaining a method of modifying a plurality of probability data according to an embodiment.
FIG. 12 is a diagram illustrating a plurality of probability data, a plurality of modified probability data, and a probability model according to an embodiment.
Figure 13 is a diagram for explaining an image decoding method according to an embodiment.
FIG. 14 is a diagram illustrating the configuration of a video encoding device according to an embodiment.
Figure 15 is a diagram illustrating the configuration of a generator according to an embodiment.
Figure 16 is a diagram for explaining the operation of the AI control unit according to one embodiment.
Figure 17 is a diagram for explaining an image encoding method according to an embodiment.
Figure 18 is a diagram for explaining a method of training neural networks according to an embodiment.

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

In describing the embodiments, if it is determined that detailed descriptions of related known technologies may unnecessarily obscure the gist of the present disclosure, the detailed descriptions may be omitted. In addition, numbers (eg, first, second, etc.) used in the description of the specification correspond to identification symbols for distinguishing one component from another component.

In the present disclosure, the expression “at least one of a, b, or c” refers to “a”, “b”, “c”, “a and b”, “a and c”, “b and c”, “a, b and c", or variations thereof.

In the present disclosure, when a component is referred to as “connected” or “connected” to another component, the component may be directly connected or directly connected to the other component, but in particular, the contrary is not specified. Unless a base material exists, it may be connected or connected through another component in the middle.

In the present disclosure, components expressed as '~unit (unit)', 'module', etc. are two or more components combined into one component, or one component divided into two or more for more detailed functions. It could be. Each of the components described below may additionally perform some or all of the functions performed by other components in addition to the main functions that each component is responsible for, and some of the main functions performed by each component may be performed by other components. It may also be carried out in full charge by .

In this disclosure, 'image' may refer to a still image, a picture, a frame, a moving image composed of a plurality of consecutive still images, or a video.

In this disclosure, 'neural network' is a representative example of an artificial neural network model that simulates brain nerves, and is not limited to an artificial neural network model using a specific algorithm. A neural network may also be referred to as a deep neural network.

In the present disclosure, a 'parameter' is a value used in the calculation process of each layer forming a neural network, and can be used, for example, when applying an input value to a predetermined calculation equation. Parameters are values set as a result of training, and can be updated through separate training data as needed.

In the present disclosure, 'feature data' may refer to data obtained by a neural network or a neural network-based encoder processing input data. Feature data may be one-dimensional or two-dimensional data including several samples. Feature data may also be referred to as a latent representation. Feature data may represent potential features in data output by a neural network-based decoder.

In this disclosure, 'current image' refers to an image that is currently the subject of processing, and 'previous image' refers to an image that is the subject of processing before the current image. The 'current video' or 'previous video' may be a block divided from the current video or the previous video.

In the present disclosure, 'sample' refers to data that is assigned to a sampling position within one-dimensional or two-dimensional data such as images, feature data, probability data, or quantization data and is subject to processing. For example, a sample may include pixels in a two-dimensional image. Two-dimensional data may also be referred to as a 'map'.

An AI-based end-to-end encoding/decoding system can be understood as a system in which a neural network is used in the video encoding and decoding process.

Like codecs such as HEVC and VVC, intra prediction or inter prediction can be used to encode and decode images in an AI-based end-to-end encoding/decoding system.

As described above, intra prediction is a method of compressing an image by removing spatial redundancy within the image, and inter prediction is a method of compressing an image by removing temporal redundancy between images.

In one embodiment, intra prediction may be applied to the first frame among several frames, a frame that serves as a random access point, and a frame in which a scene change occurs.

In one embodiment, inter prediction may be applied to frames that follow a frame to which intra prediction is applied among several frames.

With reference to FIGS. 1 and 2 , intra prediction and inter prediction performed by an AI-based end-to-end encoding/decoding system according to an embodiment will be described.

FIG. 1 is a diagram illustrating the encoding and decoding process of a current image 100 based on intra prediction according to an embodiment.

In intra prediction, the video encoder 12 and the video decoder 14 may be used. The video encoder 12 and the video decoder 14 may be implemented with a neural network.

The video encoder 12 may process the current video 100 according to parameters set through training and output feature data (k) for the current video 100.

A bitstream is generated by applying quantization 22 and entropy coding 32 to the feature data k of the current image 100, and the bitstream can be transmitted from the image encoding device to the image decoding device.

Entropy decoding 34 and inverse quantization 24 are applied to the bitstream to obtain restored feature data (k'), and the restored feature data (k') can be input to the video decoder 14.

The image decoder 14 may process the feature data (k') according to parameters set through training and output the current reconstructed image 300.

Since spatial features within the current image 100 are considered in intra prediction, only the current image 100 can be input to the image encoder 12, unlike inter prediction shown in FIG. 2.

FIG. 2 is a diagram illustrating the encoding and decoding process of the current image 100 based on inter prediction according to an embodiment.

In inter prediction, the optical flow encoder 42, optical flow decoder 44, residual encoder 52, and residual decoder 54 may be used.

The optical flow encoder 42, optical flow decoder 44, residual encoder 52, and residual decoder 54 may be implemented with a neural network.

The optical flow encoder 42 and the optical flow decoder 44 can be understood as a neural network for extracting optical flow (g) from the current image 100 and the previous reconstructed image 200.

The residual encoder 52 and the residual decoder 54 can be understood as neural networks for encoding and decoding the residual image r.

As described above, inter prediction is a process of encoding and decoding the current image 100 using temporal redundancy between the current image 100 and the previous reconstructed image 200. The previous restored image 200 may be an image obtained through decoding a previous image that was the subject of processing before processing the current image 100.

The position difference (or motion vector) between blocks or samples in the current image 100 and reference blocks or reference samples in the previous reconstructed image 200 can be used for encoding and decoding the current image 100. . This position difference can be referred to as optical flow. Optical flow may be defined as a set of motion vectors corresponding to samples or blocks in an image.

Optical flow (g) refers to how the positions of samples in the previous restored image 200 have changed in the current image 100, or samples that are the same/similar to the samples in the current image 100 in the previous restored image 200. It can indicate where it is located.

For example, if the sample that is the same or most similar to the sample located at (1, 1) in the current image 100 is located at (2, 1) in the previous restored image 200, the optical flow (g) for that sample Alternatively, the motion vector can be derived as (1(=2-1), 0(=1-1)).

To encode the current image 100, the previous reconstructed image 200 and the current image 100 may be input to the optical flow encoder 42.

The optical flow encoder 42 may process the current image 100 and the previous reconstructed image 200 according to parameters set as a result of training and output feature data (w) of the optical flow (g).

As described in FIG. 1, quantization 22 and entropy encoding 32 are applied to the feature data w of the optical flow g to generate a bitstream, and entropy decoding 34 and inverse encoding are performed on the bitstream. Quantization 24 may be applied to restore the feature data w of the optical flow g.

The feature data (w) of the optical flow (g) may be input to the optical flow decoder (44). The optical flow decoder 44 may process the input feature data (w) according to parameters set as a result of training and output an optical flow (g).

The previous restored image 200 is warped through warping 60 based on the optical flow (g), and the current predicted image (x') can be obtained as a result of the warping (60). Warping 60 is a type of geometric transformation that moves the positions of samples in an image.

Warping 60 is applied to the previous restored image 200 according to the optical flow (g), which represents the relative positional relationship between the samples in the previous restored image 200 and the samples in the current image 100, so that the current image ( A current prediction image (x') similar to 100) can be obtained.

For example, if the sample located at (1, 1) in the previous restored image 200 is most similar to the sample located at (2, 1) in the current image 100, the previous restored image ( 200) The location of the sample located at (1, 1) may be changed to (2, 1).

Since the current predicted image (x') generated from the previous restored image 200 is not the current image 100 itself, the residual image (r) between the current predicted image (x') and the current image 100 will be obtained. You can.

For example, the residual image (r) may be obtained by subtracting sample values in the current prediction image (x') from sample values in the current image (100).

The residual image r may be input to the residual encoder 52. The residual encoder 52 may process the residual image r according to parameters set as a result of training and output feature data v of the residual image r.

As described in FIG. 1, quantization 22 and entropy encoding 32 are applied to the feature data v of the residual image r to generate a bitstream, and entropy decoding 34 and inverse encoding are performed on the bitstream. Quantization 24 may be applied to restore the feature data v of the residual image r.

Feature data (v) of the residual image (r) may be input to the residual decoder 54. The residual decoder 54 may process the input feature data (v) according to parameters set as a result of training and output a restored residual image (r').

The current reconstructed image 300 can be obtained by combining the current prediction image (x') and the reconstructed residual image (r').

Meanwhile, as described above, entropy encoding 32 is performed on the feature data (k) of the current image 100, the feature data (w) of the optical flow (g), and the feature data (v) of the residual image (r). and entropy decoding 34 may be applied. Entropy coding is an encoding method that varies the average length of a code representing a symbol depending on the probability of the symbol. Therefore, in entropy coding (32) and entropy decoding (34) of the first feature data, the samples of the first feature data may have Probabilities of values may be needed.

In one embodiment, at least one of the feature data (k) of the current image 100, the feature data (w) of the optical flow (g), or the feature data (v) of the residual image (r) (hereinafter referred to as first feature data) ), probability data can be obtained based on a neural network to improve the efficiency of entropy encoding (32)/entropy decoding (34).

Probability data is one-dimensional or two-dimensional data, and a sample of the probability data may represent the probability of a value that a sample of the first feature data may have.

In one embodiment, probabilities of values that samples of the first feature data may have can be derived by applying sample values of the probability data to a predetermined probability model (e.g., a Laplacian probability model or a Gaussian probability model, etc.). there is.

In one embodiment, the probability data may include a mean and standard deviation (or variance) corresponding to a sample of the first feature data as sample values.

A method of acquiring probability data using a neural network will be described with reference to FIG. 3.

FIG. 3 is a diagram illustrating a method of obtaining probability data used for entropy encoding and entropy decoding according to an embodiment.

A hyperprior encoder 310 and a probabilistic neural network 330 may be used to obtain probability data used for entropy encoding (32)/entropy decoding (34).

The hyperfryer encoder 310 may be a neural network for obtaining feature data from feature data.

Referring to FIG. 3, first feature data is input to the hyperprior encoder 310, and the hyperprior encoder 310 processes the first feature data according to parameters set as a result of training and outputs second feature data. You can.

Since the second feature data may represent features latent in the first feature data, it may be referred to as hyperprior feature data.

The second feature data is input to the probabilistic neural network 330, and the probabilistic neural network 330 may process the second feature data according to parameters set as a result of training and output probability data.

Probability data can be used for entropy encoding 32 and entropy decoding 34 described in connection with FIGS. 1 and 2.

In one embodiment, in the encoding process of the current image 100, quantized first feature data, for example, quantized feature data of the current image 100, quantized feature data of the optical flow (g), or residual image A bitstream may be obtained by applying entropy encoding 32 based on probability data to at least one of the quantized feature data of (r).

In one embodiment, in the decoding process of the current image 100, entropy decoding 34 based on probability data is applied to the bits included in the bitstream to generate quantized first feature data, for example, the current image ( At least one of quantized feature data of 100), quantized feature data of the optical flow (g), or quantized feature data of the residual image (r) may be obtained.

The process of acquiring probability data shown in FIG. 3 may be useful when quantization (22) and dequantization (24) of the first feature data are performed uniformly. This is because the probability data obtained from the second feature data representing the latent feature in the first feature data can also be applied to the quantized first feature data.

That quantization (22) and dequantization (24) are performed uniformly may mean that the quantization step sizes used for quantization (22) and inverse quantization (24) of samples of the first feature data are all the same. For example, when the quantization step size is 2, the first feature data may be quantized by dividing all sample values of the first feature data by 2 and rounding the resulting values. Additionally, the quantized first feature data may be dequantized by multiplying all sample values of the quantized first feature data by 2.

When quantization 22 is performed based on one quantization step size, the distribution of sample values of the first feature data before quantization 22 may be maintained similarly in the first feature data after quantization 22. This is because the sample values of the first feature data are all divided and rounded based on the same value. That is, since the distribution of sample values of the first feature data before quantization (22) remains the same in the first feature data after quantization (22), the probability data obtained from the second feature data can be applied as is to the quantized first feature data. It is possible.

Uniform quantization and uniform inverse quantization may be useful when sample values of the first feature data follow a Laplacian distribution. However, since the distribution of sample values of the first feature data may vary depending on the characteristics of the image, there may be some limitations to uniform quantization and uniform inverse quantization.

In one embodiment, quantization 22 and inverse data are obtained by obtaining data related to quantization 22, for example, a quantization step size, based on a neural network from the hyperprior feature data of the first feature data that is the target of quantization 22. The efficiency of quantization (24) can be improved.

Figure 4 illustrates a method of obtaining quantization data used for quantization (22) and dequantization (24) and modified probability data used for entropy encoding (32) and entropy decoding (34) according to an embodiment. This is a drawing for

Referring to FIG. 4, first feature data is input to the hyperprior encoder 310, and the hyperprior encoder 310 processes the first feature data according to parameters set as a result of training and outputs second feature data. You can.

The second feature data may be input to the neural network 410. The neural network 410 may process the second feature data according to parameters set as a result of training and output probability data and quantization data.

Quantization data may include a quantization step size or quantization parameter as a sample value.

The quantization step size is a value used for quantization 22 of the sample, and the sample value can be quantized by dividing the sample value by the quantization step size and rounding the result of the division. Conversely, the quantized sample value may be dequantized by multiplying the quantized sample value by the quantization step size.

The quantization step size can be approximated by Equation 1 below.

[Equation 1]

Quantization step size = 2^(quantization parameter/n) / quantization scale[quantization parameter%n]

In Equation 1, quantization scale [quantization parameter %n] represents the scale value indicated by the quantization parameter among n predetermined scale values. The HEVC codec defines six scale values (26214, 23302, 20560, 18396, 16384, and 14564), so n is 6 according to the HEVC codec.

In one embodiment, when the quantization data includes a quantization parameter as a sample, the quantization step size may be obtained from the quantization parameter for quantization (22) and dequantization (24) of the first feature data. For example, Equation 1 described above can be used to derive the quantization step size.

When the quantization step size is obtained from the quantization data, samples of the first feature data may be quantized according to the quantization step size, and samples of the quantized first feature data may be dequantized according to the quantization step size.

In one embodiment, the quantization step size for samples of the first feature data is adaptively obtained for each sample from the trained neural network 410 through a training process described later with reference to FIG. 18, so uniform quantization and uniform inverse quantization It is possible to obtain a high quality current restored image (300) with a lower bit rate compared to .

The first feature data is quantized according to the quantization data obtained through the process shown in FIG. 4, so that the distribution of sample values of the first feature data before quantization (22) and the distribution of sample values of the first feature data after quantization (22) may vary. Accordingly, there may be a need to modify probability data suitable for the above-described uniform quantization and uniform inverse quantization.

In one embodiment, corrected probability data may be obtained by applying a correction process 430 based on quantization data to probability data output from the neural network 410. The modified probability data can be used for entropy encoding (32) for the quantized first feature data and entropy decoding (34) for the bitstream.

A method for modifying probability data will be described later with reference to FIGS. 9 to 12.

In the embodiment described in relation to FIG. 4, quantization data and probability data are obtained from second feature data corresponding to potential features of the first feature data that are subject to quantization 22 and entropy encoding 32, and quantization data Since the probability data is modified according to , the efficiency of quantization (22) and entropy encoding (32) performed sequentially can be improved.

FIG. 5 is a diagram illustrating the configuration of an image decoding device 500 according to an embodiment.

Referring to FIG. 5 , the image decoding apparatus 500 according to an embodiment may include an acquisition unit 510 and a prediction decoding unit 530.

The acquisition unit 510 and prediction decoding unit 530 may be implemented with at least one processor. The acquisition unit 510 and the predictive decoding unit 530 may operate according to at least one instruction stored in memory (not shown).

Figure 5 shows the acquisition unit 510 and the prediction decoding unit 530 separately, but the acquisition unit 510 and the prediction decoding unit 530 may be implemented through one processor. In this case, the acquisition unit 510 and the predictive decoding unit 530 may be implemented with a dedicated processor, or may be implemented with a general-purpose processor such as an application processor (AP), a central processing unit (CPU), or a graphic processing unit (GPU) and software. It can also be implemented through combination. Additionally, in the case of a dedicated processor, it may include a memory for implementing an embodiment of the present disclosure, or a memory processing unit for using an external memory.

The acquisition unit 510 and prediction decoding unit 530 may be implemented with a plurality of processors. In this case, the acquisition unit 510 and the prediction decoding unit 530 may be implemented through a combination of dedicated processors, or may be implemented through a combination of software and a number of general-purpose processors such as AP, CPU, or GPU.

The acquisition unit 510 may acquire a bitstream generated through neural network-based encoding of the current image 100. The bitstream may be generated through intra prediction described in relation to FIG. 1 or inter prediction described in connection with FIG. 2 .

The acquisition unit 510 may receive a bitstream from the video encoding device 1400 through a network. In one embodiment, the acquisition unit 510 is a 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. A bitstream can also be obtained from a data storage medium including -optical medium).

The acquisition unit 510 may acquire dequantized first feature data from the bitstream.

The first feature data is feature data (k) of the current video 100 output from the video encoder 12, feature data (w) of the optical flow (g) output from the optical flow encoder 42, or residual encoder 52. ) may include at least one of the feature data (v) of the residual image (r) output from.

In one embodiment, the acquisition unit 510 may obtain second feature data for the first feature data from a bitstream, and obtain modified probability data and quantization data using the second feature data. Additionally, the acquisition unit 510 may obtain the dequantized first feature data through entropy decoding and inverse quantization of the bits included in the bitstream.

The dequantized first feature data is transmitted to the prediction decoder 530, and the prediction decoder 530 can obtain the current reconstructed image 300 by applying the dequantized first feature data to a neural network. The currently restored image 300 may be output to a display device for playback.

In one embodiment, the prediction decoder 530 may obtain the current reconstructed image 300 by applying the dequantized first feature data to the image decoder 14. In this case, the prediction decoder 530 can be understood as restoring the current image 100 through intra prediction.

In one embodiment, the prediction decoder 530 applies the dequantized first feature data, for example, the dequantized feature data of the optical flow (g) to the optical flow decoder 44 to generate the optical flow (g). can be obtained. In addition, the prediction decoder 530 may obtain the restored residual image (r') by applying the dequantized feature data of the residual image (r) to the residual decoder 54. The prediction decoder 530 combines the current prediction image (x') obtained from the previous reconstructed image 200 and the reconstructed residual image (r') based on the optical flow (g) to create the current reconstructed image 300. It can be obtained. In this case, the prediction decoder 530 can be understood as restoring the current image 100 through inter prediction.

FIG. 6 is a diagram illustrating the configuration of the acquisition unit 510 according to an embodiment.

Referring to FIG. 6, the acquisition unit 510 may include an entropy decoding unit 610, an inverse quantization unit 630, and an AI control unit 650.

The bitstream is input to the entropy decoding unit 610, and the entropy decoding unit 610 can obtain quantized second feature data by applying entropy decoding to the bits included in the bitstream.

As will be described later, the second feature data may be data obtained by processing the first feature data by the hyperprior encoder 310. The image encoding apparatus 1400 may quantize the second feature data and entropy-encode the quantized second feature data to generate a bitstream including bits corresponding to the quantized second feature data.

In one embodiment, quantization may not be applied to the second feature data. In this case, the entropy decoding unit 610 may obtain second feature data by applying entropy decoding to the bits included in the bitstream, and transmit the obtained second feature data to the AI control unit 650.

The quantized second feature data may be transmitted to the inverse quantization unit 630. The inverse quantization unit 630 may inversely quantize the quantized second feature data and transmit the inverse quantized second feature data to the AI control unit 650.

In one embodiment, the quantized second feature data obtained by the entropy decoding unit 610 may be provided from the entropy decoding unit 610 to the AI control unit 650. This means that inverse quantization is skipped for the quantized second feature data.

In one embodiment, the entropy decoder 610 may use predetermined probability data to obtain second feature data (unquantized second feature data or quantized second feature data) from a bitstream.

Probability data used to obtain the second feature data may be determined based on a rule. For example, the entropy decoder 610 may determine probability data used to obtain second feature data according to predefined rules without using a neural network.

In one embodiment, the entropy decoder 610 may acquire probability data used to obtain the second feature data based on a pre-trained neural network.

In one embodiment, the inverse quantization unit 630 may use predetermined quantization data to inverse quantize the quantized second feature data.

Quantization data used to obtain the second feature data may be determined based on a rule. For example, the inverse quantization unit 630 may determine quantization data used to inverse quantize second feature data quantized according to predefined rules without using a neural network. For example, the inverse quantization unit 630 may inverse quantize the quantized second feature data according to a predetermined quantization step size.

In one embodiment, the inverse quantization unit 630 may inverse quantize sample values of the quantized second feature data according to the same quantization step size.

In one embodiment, the inverse quantization unit 630 may obtain quantization data used to inverse quantize the quantized second feature data based on a pre-trained neural network.

The AI control unit 650 may obtain modified probability data and quantized data using second feature data, specifically unquantized second feature data, quantized second feature data, or dequantized second feature data. .

In one embodiment, the AI controller 650 may use a neural network to obtain modified probability data and quantization data.

The modified probability data may be transmitted to the entropy decoding unit 610, and the quantized data may be transmitted to the inverse quantization unit 630.

The entropy decoder 610 may obtain quantized first feature data by applying entropy decoding based on modified probability data to the bits included in the bitstream. The quantized first feature data may be transmitted to the inverse quantization unit 630.

The inverse quantization unit 630 may inverse quantize the quantized first feature data based on the quantization data transmitted from the AI control unit 650 and transmit the inverse quantized first feature data to the prediction decoder 530. .

Referring to FIG. 7, the operation of the AI control unit 650 will be described in more detail.

FIG. 7 is a diagram for explaining the operation of the AI control unit 650 according to an embodiment.

The AI control unit 650 may use the probabilistic neural network 330 and the quantization neural network 350 to obtain probability data and quantization data.

The probabilistic neural network 330 and the quantization neural network 350 may be stored in memory. In one embodiment, the probabilistic neural network 330 and the quantization neural network 350 may be implemented with an AI processor.

Second feature data (specifically, unquantized second feature data, quantized second feature data, or dequantized second feature data) may be input to the probabilistic neural network 330 and the quantization neural network 350.

The probability neural network 330 may process the second feature data according to parameters set as a result of training and output probability data.

The quantization neural network 350 may process the second feature data according to parameters set as a result of training and output quantization data.

As previously described in relation to FIG. 4 , probability data and quantization data may be output by processing the second feature data by one neural network 410.

Quantization data may include a quantization parameter or a quantization step size, and probability data may include values representing probabilities of values that samples of the first feature data may have. In one embodiment, the probability data may include a mean, standard deviation, and/or variance for each sample of the first feature data.

In one embodiment, the size or number of samples of the probability data and quantization data may be the same as the size or number of samples of the first feature data.

As described above, since the distribution of sample values of the first feature data can be changed through adaptive quantization based on quantization data, the probability data can be modified through a correction process 400 based on quantization data.

In one embodiment, the AI control unit 650 may obtain modified probability data by dividing sample values of probability data by sample values of quantization data. Here, division is an example, and in one embodiment, the AI control unit 650 multiplies the sample values of the probability data by the sample values of the quantized data or the values derived from the sample values of the quantized data to obtain modified probability data. It may be possible.

In one embodiment, the AI control unit 650 may use a bit shift operation for a division operation or multiplication operation on sample values of probability data.

In one embodiment, the modification process 430 may be performed based on a neural network. For example, corrected probability data may be obtained by applying the probability data and quantization data to a neural network for the correction process 430.

The AI control unit 650 may transmit the corrected probability data to the entropy decoding unit 610 and transfer the quantized data to the inverse quantization unit 630.

The entropy decoding unit 610 may obtain quantized first feature data by applying entropy decoding based on modified probability data to the bits of the bitstream. The inverse quantization unit 630 may obtain inverse quantized first feature data by inverse quantizing the quantized first feature data according to the quantization data.

Exemplary structures of the probabilistic neural network 330, the quantization neural network 350, the neural network for the correction process shown in FIG. 7, and the hyperprior encoder 310 described above will be described with reference to FIG. 8.

FIG. 8 is a diagram illustrating the structure of a neural network 800 according to one embodiment.

As shown in FIG. 8, input data 805 may be input to the first convolution layer 810. 3X3X4 displayed on the first convolution layer 810 illustrates convolution processing on the input data 805 using four filter kernels of size 3x3. As a result of convolution processing, four feature data can be generated by four filter kernels.

In one embodiment, if the neural network 800 corresponds to a neural network for a correction process, the input data 805 may include two channels of data, that is, probability data and quantization data.

In one embodiment, if the neural network 800 corresponds to the probabilistic neural network 330 or the quantization neural network 350, the input data 805 may include second feature data.

In one embodiment, if the neural network 800 corresponds to the hyperprior encoder 310, the input data 805 may include first feature data.

Feature data generated by the first convolutional layer 810 may represent unique characteristics of the input data 805. For example, each feature data may represent vertical characteristics, horizontal characteristics, or edge characteristics of the input data 805.

Feature data of the first convolutional layer 810 may be input to the first activation layer 820.

The first activation layer 820 may provide non-linear characteristics to each feature data. The first activation layer 820 may include a sigmoid function, a Tanh function, a Rectified Linear Unit (ReLU) function, etc., but is not limited thereto.

Giving nonlinear characteristics in the first activation layer 820 may mean changing and outputting some sample values of feature data. At this time, the change can be performed by applying non-linear characteristics.

The first activation layer 820 may determine whether to transmit sample values of feature data to the second convolution layer 830. For example, among the sample values of feature data, some sample values are activated by the first activation layer 820 and transmitted to the second convolution layer 830, and some sample values are activated by the first activation layer 820. It may be deactivated and not transmitted to the second convolution layer 830. Unique characteristics of the input data 805 indicated by the feature data may be emphasized by the first activation layer 820.

Feature data output from the first activation layer 820 may be input to the second convolution layer 830. 3X3X4 displayed on the second convolution layer 830 illustrates convolution processing on input feature data using four filter kernels of size 3x3.

The output of the second convolution layer 830 may be input to the second activation layer 840. The second activation layer 840 may provide non-linear characteristics to the input feature data.

Feature data output from the second activation layer 840 may be input to the third convolution layer 850. 3

The output data 855 varies depending on whether the neural network 800 is a hyperprior encoder 310, a probabilistic neural network 330, a quantization neural network 350, or a neural network for the correction process.

For example, if the neural network 800 is a probabilistic neural network 330, the output data 855 may be probability data, and if the neural network 800 is a quantization neural network 350, the output data 855 may be quantized data. .

In one embodiment, the number of output data 855 may be adjusted by adjusting the number of filter kernels used in the third convolution layer 850.

For example, if the neural network 800 is a probabilistic neural network 330 and the probability data includes average data and standard deviation data described later, two channels of data may be output from the third convolution layer 850. Two filter kernels can be used so that

Additionally, for example, if the neural network 800 is a probabilistic neural network 330, the probability data includes average data and standard deviation data described later, and the number of channels of the first feature data is M, then the M average data and M 2M filter kernels may be used in the third convolution layer 850 so that standard deviation data can be output.

In addition, for example, if the neural network 800 is a quantization neural network 350 and the number of channels of the first feature data is M, M filter kernels are used in the third convolution layer 850 so that M quantized data can be output. ) can be used in.

Figure 8 shows that the neural network 800 has three convolutional layers (first convolutional layer 810, second convolutional layer 830, and third convolutional layer 850) and two activation layers (first activation layer). Although it is shown as including a layer 820 and a second activation layer 840, this is only an example. In one embodiment, the number of convolution layers and activation layers included in the neural network 800 is It can be changed in various ways.

In one embodiment, the size and number of filter kernels used in convolutional layers included in the neural network 800 may also vary.

In one embodiment, the neural network 800 may be implemented through a recurrent neural network (RNN). This means changing the CNN structure of the neural network 800 to an RNN structure.

In one embodiment, the image decoding device 500 and the image encoding device 1400 may include at least one Arithmetic logic unit (ALU) for convolution operation and activation layer operation.

ALU can be implemented as a processor. For the convolution operation, the ALU may include a multiplier that performs a multiplication operation between sample values of input data or feature data output from the previous layer and sample values of the filter kernel, and an adder that adds the resultant values of the multiplication.

For the operation of the activation layer, the ALU has a multiplier that multiplies the input sample value by a weight used in a predetermined sigmoid function, Tanh function, or ReLU function, and compares the multiplication result with a predetermined value to transfer the input sample value to the next layer. It may include a comparator that determines whether to forward it to .

Hereinafter, with reference to FIGS. 9 to 12, probability data used for entropy encoding and entropy decoding will be described.

Figure 9 is a diagram for explaining a method of modifying probability data according to an embodiment.

In one embodiment, probability data output by the probabilistic neural network 330 may represent probabilities of values that samples of the first feature data may have.

In one embodiment, the probability data may include averages and standard deviations corresponding to samples of the first feature data. In this case, the probability data includes average data 910 containing as samples averages corresponding to samples of the first feature data and standard deviation data 930 including standard deviations corresponding to samples of the first feature data as samples. ) may include.

In one embodiment, the probabilistic neural network 330 includes mean data whose samples include means corresponding to samples of the first feature data and variance data whose samples include variances corresponding to samples of the first feature data. can do.

In one embodiment, the AI control unit 650 may obtain modified probability data by dividing sample values of probability data by sample values of quantization data.

Referring to FIG. 9, average data 910 includes sample values from μ(0,0) to μ(1,1), and standard deviation data 930 includes sample values from σ(0,0) to σ(1, 1), and the quantization data 950 may include sample values of q(0,0) to q(1,1). q(0,0) to q(1,1) of the quantization data 950 may be quantization step sizes. In Figure 9, (a, b) may indicate the location of the sample within the data.

The AI control unit 650 divides μ(0,0) to μ(1,1) in the average data 910 by q(0,0) to q(1,1) in the quantized data 950, Corrected average data 970 containing samples from (0,0)/q(0,0) to μ(1,1)/q(1,1) can be obtained.

In addition, the AI control unit 650 divides σ(0,0) to σ(1,1) in the standard deviation data 930 by q(0,0) to q(1,1) in the quantization data 950. As a result, modified standard deviation data 990 including σ(0,0)/q(0,0) to σ(1,1)/q(1,1) as samples can be obtained.

Here, division is an example, and in one embodiment, the AI control unit 650 combines the sample values of the average data 910 and the standard deviation data 930 with the sample values of the quantization data 950 or the quantization data 950. The values derived from the sample values of may be multiplied to obtain the corrected average data 970 and the corrected standard deviation data 990.

In one embodiment, the AI control unit 650 may use a bit shift operation to perform a division or multiplication operation on sample values of the average data 910 and the standard deviation data 930.

The AI control unit 650 applies the sample values of the corrected average data 970 and the sample values of the corrected standard deviation data 990 to a predetermined probability model to generate the first feature data quantized according to the quantization data 950. Probability values that samples may have can be derived.

To explain the reason for dividing the sample values of the average data 910 and the sample values of the standard deviation data 930 into the sample values of the quantization data 950, the sample value of the first feature data based on the quantization data 950 is When the resulting value is divided and rounded, the sample value of the first feature data becomes larger or smaller depending on the size of the quantization step size, and accordingly, a probability model (e.g., probability density function) for the first feature data Because this needs to change. Accordingly, a probability model suitable for the quantized first feature data can be derived by downscaling the average data 910 and the standard deviation data 930 according to the quantization data 950.

Figure 10 is a diagram for explaining probability data, modified probability data, and probability model according to an embodiment.

Referring to FIG. 10, the probability data may include a mean (μ _a ) and a standard deviation (σ _a ), and the mean (μ _a ) and standard deviation (σ _a ) of the probability data are the quantization step size of the quantization data ( As divided by q), the adjusted probability data may include a modified mean (μ _b ) and a modified standard deviation (σ _b ).

The AI control unit 650 may determine the probability of a value that a sample of the quantized first feature data may have by applying the modified mean (μ _b ) and the modified standard deviation (σ _b ) to a predetermined probability model.

Referring to FIG. 10, a Laplacian probability model or a Gaussian probability model may be used as a predetermined probability model.

The Laplacian probability model or Gaussian probability model shown in FIG. 10 is an example. In one embodiment, the types of probability models used for entropy encoding and entropy decoding may vary.

The probability model to be used for entropy encoding and entropy decoding of the first feature data may be determined in advance. For example, the type of probability model used for entropy encoding may be determined in advance by the image encoding apparatus 1400.

In one embodiment, the type of probability model used for entropy encoding may be independently determined for each image or for each block included in the image.

When the Laplacian model is used for entropy encoding, the AI control unit 650 applies the modified mean (μ _b ) and the modified standard deviation (σ _b ) to the Laplacian probability model to obtain a sample of the quantized first feature data. The probability can be derived.

In addition, when a Gaussian model is used for entropy encoding, the AI control unit 650 applies the modified mean (μ _b ) and the modified standard deviation (σ _b ) to the Gaussian probability model to obtain a sample of the quantized first feature data. You can derive the probability that you can have.

In one embodiment, the probabilistic neural network 330 may output a plurality of probability data and a plurality of weight data as a result of processing the second feature data.

In one embodiment, each of the plurality of probability data may include average data and standard deviation data. In one embodiment, each of the plurality of probability data may include average data and variance data.

The AI control unit 650 may obtain a plurality of corrected probability data by modifying the plurality of probability data according to the quantization data.

As a plurality of modified probability data are combined according to a plurality of weight data, the probability that a sample of the quantized first feature data may have can be derived.

Figure 11 is a diagram for explaining a method of modifying a plurality of probability data according to an embodiment.

Referring to FIG. 11, N average data 910-1, 910-2, ..., 910-N can be obtained from the probabilistic neural network 330.

Although not shown in FIG. 11, N standard deviation data and N weight data may also be obtained from the probabilistic neural network 330.

The size or number of samples of N average data (910-1, 910-2, ..., 910-N), N standard deviation data, and N weight data will be the same as the size or number of samples of the first feature data. You can.

In one embodiment, if the number of first feature data (or number of channels) is M, M*N average data, M*N standard deviation data, and M*N weight data are obtained from the probabilistic neural network 330. It can be.

The AI control unit 650 divides the sample values of the N average data (910-1, 910-2, ..., 910-N) by the sample values of the quantization data (950) to produce the N modified average data (970-N). 1, 970-2, ..., 970-N) can be obtained.

The AI control unit 650 may obtain N corrected standard deviation data by dividing the sample values of the N standard deviation data by the sample values of the quantization data 950.

Figure 12 is a diagram for explaining probability data, modified probability data, and probability model according to an embodiment.

Referring to FIG. 12, when N average data (μ _a ), N standard deviation data (σ _a ), and N weight data (w) are obtained from the probabilistic neural network 330, the AI control unit 650 Divide the average data (μ _a ) and N standard deviation data (σ _a ) by the sample values of the quantization data (q) to obtain N corrected average data (μ _b ) and N corrected standard deviation data (σ _b ). can be obtained.

In one embodiment, the N average data (μ _a ) shown in FIG. 12 may include the N average data (910-1, 910-2, ..., 910-N) shown in FIG. 11. . For example, μ _a1 in FIG. 12 may correspond to the average data 910-1 in FIG. 11. Also, for example, μ _b1 in FIG. 12 may correspond to modified average data 970-1, which is the result of dividing μ _a1 by quantization data 950.

The AI control unit 650 converts N corrected average data (μ _b ), N corrected standard deviation data (σ _b ), and N weight data (w) into a predetermined probability model, for example, shown in FIG. 12 By applying the Laplacian probability model or the Gaussian probability model, the probabilities of values that samples of the quantized first feature data may have can be derived.

According to one embodiment, N corrected average data (μ _b ) and N corrected standard deviation data (σ _b ) are acquired for one first feature data, and N corrected average data (μ _b ) and N corrected standard deviation data (σ _b ) are combined according to N weight data (w), so the probability that a sample of the first feature data may have can be derived more accurately and stably.

Figure 13 is a diagram for explaining an image decoding method according to an embodiment.

In step S1310, the image decoding device 500 may obtain second feature data for the first feature data obtained through neural network-based encoding of the current image 100 from a bitstream.

In one embodiment, the first feature data is feature data (k) obtained by applying the current image 100 to the image encoder 12, and the current image 100 and the previous reconstructed image 200 are applied to the optical flow encoder 42. ) or feature data (v) obtained by applying the residual image (r) corresponding to the current image 100 to the residual encoder 52.

In one embodiment, the image decoding apparatus 500 may obtain second feature data by applying entropy decoding to bits included in the bitstream.

In one embodiment, the image decoding apparatus 500 may obtain quantized second feature data by applying entropy decoding to bits included in the bitstream, and dequantize the quantized second feature data.

In step S1320, the image decoding device 500 applies the second feature data to the neural network to obtain quantization data and probability data.

In one embodiment, the neural network may include a probabilistic neural network 330 that outputs probability data and a quantization neural network 350 that outputs quantized data.

In step S1330, the image decoding device 500 modifies probability data based on sample values of quantized data.

In one embodiment, the image decoding apparatus 500 may divide sample values of probability data into sample values of quantization data. In this case, sample values of quantization data may be quantization step sizes. If the sample values of the quantization data correspond to the quantization parameters, the image decoding apparatus 500 may determine quantization step sizes from the quantization parameters and divide the sample values of the probability data by the determined quantization step sizes.

In step S1340, the image decoding apparatus 500 obtains quantized first feature data by applying entropy decoding based on modified probability data to the bits included in the bitstream.

In step S1350, the image decoding apparatus 500 dequantizes the quantized first feature data according to sample values of the quantized data to obtain dequantized first feature data.

In step S1360, the image decoding device 500 restores the current image 100 through neural network-based decoding of the dequantized first feature data.

In one embodiment, the image decoding device 500 restores the current image 100 by applying the dequantized first feature data to the image decoder 14, the optical flow decoder 44, and/or the residual decoder 54. can do.

FIG. 14 is a diagram illustrating the configuration of a video encoding device 1400 according to an embodiment.

Referring to FIG. 14 , the image encoding device 1400 may include a prediction encoding unit 1410, a generation unit 1420, an acquisition unit 1430, and a prediction decoding unit 1440.

The prediction encoding unit 1410, generation unit 1420, acquisition unit 1430, and prediction decoding unit 1440 may be implemented with a processor. The prediction encoder 1410, generation unit 1420, acquisition unit 1430, and prediction decoder 1440 may operate according to instructions stored in memory (not shown).

Figure 14 shows the prediction encoding unit 1410, the generation unit 1420, the acquisition unit 1430, and the prediction decoding unit 1440 individually, but the prediction encoding unit 1410, the generation unit 1420, and the acquisition unit 1430 and the prediction decoder 1440 may be implemented through one processor. In this case, the prediction encoder 1410, the generator 1420, the acquisition unit 1430, and the prediction decoder 1440 are implemented with a dedicated processor, an application processor (AP), a central processing unit (CPU), or a GPU ( It can also be implemented through a combination of a general-purpose processor, such as a graphic processing unit, and software. Additionally, in the case of a dedicated processor, it may include a memory for implementing an embodiment of the present disclosure, or a memory processing unit for using an external memory.

The prediction encoder 1410, generation unit 1420, acquisition unit 1430, and prediction decoder 1440 may be implemented with a plurality of processors. In this case, the prediction encoding unit 1410, generation unit 1420, acquisition unit 1430, and prediction decoding unit 1440 are implemented as a combination of dedicated processors or a plurality of general-purpose processors such as AP, CPU, or GPU. It can also be implemented through a combination of software.

The prediction encoder 1410 may obtain first feature data by applying neural network-based encoding to the current image 100. The first feature data may include at least one of feature data (k) of the current image 100, feature data (w) of the optical flow (g), or feature data (v) of the residual image (r).

In one embodiment, the prediction encoder 1410 may obtain feature data (k) of the current image 100 by applying the current image 100 to the image encoder 12.

In one embodiment, the prediction encoder 1410 may obtain feature data (w) of the optical flow (g) by applying the current image 100 and the previous reconstructed image 200 to the optical flow encoder 42. .

In one embodiment, the prediction encoder 1410 applies the residual image (r) corresponding to the difference between the current prediction image (x') and the current image 100 to the residual encoder 52 to produce a residual image (r). Characteristic data (v) can be obtained.

The first feature data obtained by the predictive encoding unit 1410 may be transmitted to the generating unit 1420.

The generator 1420 may generate a bitstream based on the first feature data.

In one embodiment, the generator 1420 may obtain second feature data representing potential features of the first feature data, and apply the second feature data to a neural network to obtain quantization data and modified probability data. Additionally, the generator 1420 may quantize the first feature data according to quantization data and entropy encode the quantized first feature data according to the modified probability data to generate a bitstream.

In one embodiment, the generator 1420 may generate a bitstream by entropy encoding the second feature data or the quantized second feature data according to predetermined probability data.

In one embodiment, the bitstream may include bits corresponding to quantized first feature data and second feature data or bits corresponding to quantized second feature data.

The bitstream may be transmitted from the video decoding device 500 through a network. In one embodiment, the bitstream is a 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. ) may also be recorded on a data storage medium containing, etc.

The acquisition unit 1430 may acquire dequantized first feature data from the bitstream generated by the generation unit 1420.

The dequantized first feature data may be transmitted to the prediction decoder 1440.

The prediction decoder 1440 may obtain the current reconstructed image 300 by applying neural network-based decoding to the dequantized first feature data.

The configuration and operation of the acquisition unit 1430 and the prediction decoding unit 1440 may be the same as those of the acquisition unit 510 and the prediction decoding unit 530 of the image decoding device 500.

FIG. 15 is a diagram illustrating the configuration of the generator 1420 according to an embodiment.

Referring to FIG. 15, the generator 1420 may include an AI control unit 1510, a quantization unit 1530, and an entropy encoding unit 1550.

The first feature data obtained by the prediction encoding unit 1410 may be transmitted to the AI control unit 1510 and the quantization unit 1530.

The AI control unit 1510 may obtain second feature data from first feature data, and obtain quantization data and modified probability data based on the second feature data. The second feature data and quantization data may be transmitted to the quantization unit 1530, and the modified probability data may be transmitted to the entropy encoding unit 1550.

In one embodiment, the AI control unit 1510 may obtain quantization data and modified probability data based on second feature data to which quantization and dequantization have been applied. Predetermined quantization data, for example, quantization data determined based on a rule, may be used to quantize and dequantize the second feature data.

The quantization unit 1530 may obtain quantized first feature data by quantizing the first feature data according to quantization data generated based on a neural network.

In one embodiment, the quantization unit 1530 may obtain quantized second feature data by quantizing the second feature data according to quantization data generated based on a rule. In one embodiment, the second feature data may be transmitted from the AI control unit 1510 to the entropy encoding unit 1550. This means that quantization for the second feature data is skipped.

The quantization unit 1530 may transmit the quantized first feature data and the quantized second feature data to the entropy encoding unit 1550.

The entropy encoding unit 1550 may generate a bitstream by entropy encoding the quantized first feature data according to modified probability data.

In one embodiment, the entropy encoding unit 1550 may generate a bitstream by entropy encoding the second feature data or the quantized second feature data according to predetermined probability data.

The bitstream may include bits corresponding to quantized first feature data and second feature data or bits corresponding to quantized second feature data.

Hereinafter, the operation of the AI control unit 1510 will be described in more detail with reference to FIG. 16.

Figure 16 is a diagram for explaining the operation of the AI control unit 1510 according to an embodiment.

The AI control unit 1510 may use a hyperprior encoder 310, a quantization neural network 350, and a probabilistic neural network 330.

The hyperprior encoder 310, quantization neural network 350, and probabilistic neural network 330 may be stored in memory. In one embodiment, hyperprior encoder 310, quantization neural network 350, and probabilistic neural network 330 may be implemented with an AI processor.

Referring to FIG. 16, first feature data may be input to the hyperfryer encoder 310.

The hyperprior encoder 310 may obtain second feature data by processing the first feature data according to parameters set as a result of training. The second feature data may be input to the probabilistic neural network 330 and the quantization neural network 350.

In one embodiment, second feature data to which quantization and dequantization have been applied may be input to the probabilistic neural network 330 and the quantization neural network 350. The reason for quantizing and dequantizing the second feature data is to consider the case where the bitstream transmitted to the image decoding device 500 includes quantized second feature data. In other words, since the image decoding device 500 can obtain probability data and quantization data using the inverse quantized second feature data, the image encoding device 1400 also performs quantization and inverse data in the same way as the image decoding device 500. The second feature data to which quantization has been applied is used.

The probability neural network 330 may process the second feature data according to parameters set as a result of training and output probability data.

The quantization neural network 350 may process the second feature data according to parameters set as a result of training and output quantization data.

As previously described in relation to FIG. 4 , probability data and quantization data may be output by processing the second feature data by one neural network 410.

Quantization data may include a quantization parameter or a quantization step size, and probability data may include values representing probabilities of values that samples of the first feature data may have. In one embodiment, the probability data may include a mean, standard deviation, and/or variance for each sample of the first feature data.

In one embodiment, the size or number of samples of the probability data and quantization data may be the same as the size or number of samples of the first feature data.

In one embodiment, probability data may be modified through a modification process 430 based on quantization data.

In one embodiment, the AI control unit 1510 may obtain modified probability data by dividing the sample values of the probability data by the sample values of the quantization data. Here, division is an example, and in one embodiment, the AI control unit 1510 multiplies the sample values of the probability data by the sample values of the quantized data or the values derived from the sample values of the quantized data to obtain modified probability data. It may be possible.

In one embodiment, the AI control unit 1510 may use a bit shift operation for a division operation or multiplication operation on sample values of probability data.

In one embodiment, the modification process 430 may be performed based on a neural network. For example, corrected probability data may be obtained by applying the probability data and quantization data to a neural network for the correction process 430.

The AI control unit 1510 may transmit the modified probability data to the entropy encoding unit 1550 and transmit the quantized data to the quantization unit 1530.

The quantization unit 1530 may obtain quantized first feature data by quantizing the first feature data according to the quantization data. Additionally, the quantization unit 1530 may obtain second feature data by quantizing the second feature data.

In one embodiment, the quantization unit 1530 may use predetermined quantization data to quantize the second feature data. Quantization data used to quantize the second feature data may be determined based on a rule. In other words, the quantization unit 1530 may determine quantization data used to quantize the second feature data according to predefined rules without using a neural network. For example, the quantization unit 1530 may quantize the second feature data according to a predetermined quantization step size. In one embodiment, the quantization unit 1530 may quantize sample values of the second feature data according to the same quantization step size.

The entropy encoding unit 1550 may generate a bitstream by applying entropy encoding based on modified probability data to the quantized first feature data. Additionally, the entropy encoding unit 1550 may entropy encode the second feature data or quantized second feature data.

In one embodiment, the entropy encoding unit 1550 may use predetermined probability data to apply entropy encoding to the second feature data or quantized second feature data. Probability data used to entropy encode the second feature data or the quantized second feature data may be determined based on a rule. In other words, the entropy encoding unit 1550 may determine probability data used to entropy encode the second feature data or quantized second feature data according to predefined rules without using a neural network.

Figure 17 is a diagram for explaining an image encoding method according to an embodiment.

In step S1710, the image encoding device 1400 applies the first feature data obtained through neural network-based encoding of the current image 100 to the hyperprior encoder 310 to generate second feature data for the first feature data. obtain.

In step S1720, the image encoding device 1400 obtains quantization data and probability data by applying the second feature data to a neural network (eg, the probability neural network 330 and the quantization neural network 350).

In one embodiment, the image encoding apparatus 1400 may obtain quantization data and probability data by applying quantized and dequantized second feature data to a neural network.

In step S1730, the image encoding device 1400 modifies probability data based on sample values of quantized data.

In one embodiment, the image encoding apparatus 1400 may divide sample values of probability data into sample values of quantization data. In this case, sample values of quantization data may be quantization step sizes. If the sample values of the quantization data correspond to the quantization parameters, the image encoding device 1400 may determine quantization step sizes from the quantization parameters and divide the sample values of the probability data by the determined quantization step sizes.

In step S1740, the image encoding device 1400 may obtain quantized first feature data by quantizing the first feature data according to sample values of the quantized data.

In one embodiment, the image encoding apparatus 1400 may obtain quantized second feature data by quantizing the second feature data according to predetermined quantization data.

In step S1750, the image encoding device 1400 may apply entropy coding based on modified probability data to the quantized first feature data to generate a bitstream including bits corresponding to the quantized first feature data. .

In one embodiment, the image encoding apparatus 1400 may entropy encode unquantized second feature data or quantized second feature data according to predetermined probability data. In this case, the bitstream may include bits corresponding to quantized first feature data, and bits corresponding to unquantized second feature data or quantized second feature data.

Hereinafter, with reference to FIG. 18, a method of training the above-described neural networks, the hyperprior encoder 310, the probabilistic neural network 330, and the quantization neural network 350 will be described.

Figure 18 is a diagram for explaining a method of training neural networks according to an embodiment.

Referring to FIG. 18, the current training image may correspond to the current image 100 described above, and the current reconstructed training image may correspond to the current reconstructed image 300 described above.

In the training process according to one embodiment, neural networks may be trained so that the current reconstructed training image becomes as similar as possible to the current training image and the bit rate of the bitstream generated through encoding of the current training image is minimized. To this end, as shown in FIG. 18, first loss information 1880 and second loss information 1890 can be used to train neural networks.

To describe the training process of neural networks in detail, first, first feature data can be obtained by applying a neural network-based encoding process (1810) to the current training image.

The neural network-based encoding process 1810 may be a process of encoding the current training image based on the image encoder 12, the optical flow encoder 42, and/or the residual encoder 52.

The first feature data is feature data obtained by processing the current training image by the image encoder 12, feature data obtained by processing the current training image and the previous reconstructed training image by the optical flow encoder 42, or current training. The residual training image corresponding to the difference between the image and the current prediction training image may include at least one of feature data obtained by being processed by the residual encoder 52. Here, the current prediction training image can be obtained by transforming the previous reconstruction training image according to the optical flow (g).

The first feature data may be input to the hyperfryer encoder 310. The hyperfryer encoder 310 may process the first feature data according to preset parameters and output second feature data.

The second feature data may be input to the probabilistic neural network 330 and the quantization neural network 350. Each of the probability neural network 330 and the quantization neural network 350 may process the second feature data according to preset parameters and output probability data and quantization data.

Probability data may be modified through a correction process 1820 according to quantization data. Since the correction process 1820 has been described in FIGS. 9 to 12, detailed description will be omitted.

A quantization process 1830 according to quantization data may be applied to the first feature data to obtain quantized first feature data. Additionally, a bitstream can be generated by applying an entropy encoding process 1840 based on modified probability data to the quantized first feature data.

In one embodiment, the bitstream may include bits corresponding to second feature data or quantized second feature data.

An entropy decoding process (1850) based on modified probability data is performed on the bitstream to obtain quantized first feature data, and an inverse quantization process (1860) according to the quantized data is performed on the quantized first feature data. Inverse quantized first feature data may be obtained.

The current reconstructed training image can be obtained by processing the dequantized first feature data according to a neural network-based decoding process (1870).

The neural network-based decoding process may be a process of restoring the current image 100 based on the image decoder 14, optical flow decoder 44, and/or residual decoder 54.

For training the neural network used in the neural network-based encoding process, the hyperprior encoder 310, the probabilistic neural network 330, the quantization neural network 350, and the neural network used in the neural network-based decoding process, first loss information 1880 Alternatively, at least one of the second loss information 1890 may be obtained.

The first loss information 1880 can be calculated from the bitrate of the bitstream generated as a result of encoding the current training image.

Since the first loss information 1880 is related to the coding efficiency for the current training image, the first loss information may be referred to as compression loss information.

The second loss information 1890 may correspond to the difference between the current training image and the current restored training image. In one embodiment, the difference between the current training image and the current reconstruction training image is the L1-norm value, L2-norm value, Structural Similarity (SSIM) value, PSNR-HVS between the current training image and the current reconstruction training image. It may include at least one of a (Peak Signal-To-Noise Ratio-Human Vision System) value, a Multiscale SSIM (MS-SSIM) value, a Variance Inflation Factor (VIF) value, or a Video Multimethod Assessment Fusion (VMAF) value.

Since the second loss information 1890 is related to the quality of the current restoration training image, it may also be referred to as quality loss information.

The neural network used in the neural network-based encoding process, the hyperprior encoder 310, the probabilistic neural network 330, the quantization neural network 350, and the neural network used in the neural network-based decoding process include first loss information (1880) or second loss. The final loss information derived from at least one of the information 1890 may be trained to be reduced or minimized.

In one embodiment, the neural network used in the neural network-based encoding process, the hyperprior encoder 310, the probabilistic neural network 330, the quantization neural network 350, and the neural network used in the neural network-based decoding process use values of preset parameters. By making changes, the final loss information can be reduced or minimized.

In one embodiment, final loss information can be calculated according to Equation 2 below.

[Equation 2]

Final loss information = a*first loss information+b*second loss information

In Equation 2, a and b are weights applied to the first loss information 1880 and the second loss information 1890, respectively.

According to Equation 2, the neural network used in the neural network-based encoding process, the hyperprior encoder 310, the probability neural network 330, the quantization neural network 350, and the neural network used in the neural network-based decoding process are the current restoration training image. It can be seen that the training is done in a way that is as similar as possible to the current training image and minimizes the size of the bitstream.

The training process described with reference to FIG. 18 may be performed by a training device. The training device may be, for example, the video encoding device 1400 or a separate server. Parameters obtained as a result of training may be stored in the video encoding device 1400 and the video decoding device 500.

The image encoding device 1400, the image decoding device 500, and the method thereof according to an embodiment have the task of efficiently quantizing/dequantizing feature data generated through AI-based encoding of an image.

The video encoding device 1400, the video decoding device 500, and the method using them according to an embodiment have the task of reducing the bit rate of the bitstream generated as a result of video encoding and improving the quality of the restored video. Do this.

The image encoding device 1400, the image decoding device 500, and the method thereof according to an embodiment have the task of improving the efficiency of entropy coding.

The video encoding device 1400, the video decoding device 500, and the method using them according to one embodiment have the task of providing an AI-based end-to-end encoding/decoding system.

A method for decoding an image according to an embodiment may include obtaining second feature data for first feature data obtained through neural network-based encoding for the current image 100 from a bitstream (S1310). there is.

In one embodiment, the image decoding method may include a step (S1320) of obtaining quantization data and probability data by applying the second feature data to the neural network (330; 350; 410).

In one embodiment, the image decoding method may include modifying probability data based on sample values of quantized data (S1330).

In one embodiment, the image decoding method may include obtaining quantized first feature data by applying entropy decoding based on modified probability data to bits included in the bitstream (S1340).

In one embodiment, the image decoding method may include the step of inversely quantizing quantized first feature data according to sample values of quantized data to obtain dequantized first feature data (S1350).

In one embodiment, the image decoding method may include restoring the current image through neural network-based decoding of the dequantized first feature data (S1360).

In one embodiment, quantization data may include quantization parameters or quantization step sizes as samples.

In one embodiment, the sample value of the modified probability data may represent the probability of a value that a sample of the quantized first feature data may have.

In one embodiment, the sample values of the modified probability data may represent the mean and standard deviation corresponding to the samples of the quantized first feature data.

In one embodiment, the probability of a value that a sample of the quantized first feature data may have can be derived by applying the average and standard deviation indicated by the sample value of the modified probability data to a predetermined probability model.

In one embodiment, modifying the probability data may include dividing sample values of the probability data by sample values of the quantization data.

In one embodiment, the first feature data is feature data (k) obtained by applying the current image 100 to the image encoder 12, and the current image 100 and the previous reconstructed image 200 are applied to the optical flow encoder ( 42) may include feature data (w) obtained by applying the residual image (r) corresponding to the current image 100 to the residual encoder 52.

In one embodiment, the neural network 330;350;410 may be trained based on at least one of the bitrate of the bitstream or the difference between the current training image and the current reconstructed training image.

In one embodiment, the second feature data is applied to the neural network (330; 410), so that a plurality of probability data and a plurality of weights are obtained, the plurality of probability data is modified based on the sample values of the quantization data, and the plurality of probability data is obtained. The probability of a value that a sample of the quantized first feature data can have can be determined by combining the modified probability data according to a plurality of weights.

An image encoding method according to an embodiment applies first feature data obtained through neural network-based encoding of the current image 100 to the first neural network 310 to generate second feature data for the first feature data. It may include a step of acquiring (S1710).

In one embodiment, the image encoding method may include a step (S1720) of obtaining quantization data and probability data by applying second feature data to a second neural network (330; 350; 410).

In one embodiment, the video encoding method may include modifying probability data based on sample values of quantized data (S1730).

In one embodiment, the image encoding method may include quantizing the first feature data according to sample values of the quantized data to obtain quantized first feature data (S1740).

In one embodiment, the image encoding method includes generating a bitstream including bits corresponding to the quantized first feature data by applying entropy coding based on modified probability data to the quantized first feature data ( S1750) may be included.

In one embodiment, the bitstream may include bits corresponding to second characteristic data.

In one embodiment, the second neural network (330;350;410) includes a quantization neural network (350) that outputs quantization data from second feature data; and a probabilistic neural network 330 that outputs probability data from the second feature data.

In one embodiment, modifying the probability data may include dividing sample values of the probability data by sample values of the quantization data.

The image decoding apparatus 500 according to an embodiment includes an acquisition unit 510 that acquires second feature data for first feature data obtained through neural network-based encoding of the current image 100 from a bitstream. may include.

In one embodiment, the image decoding apparatus 500 may include a prediction decoder 530 that restores the current image 100 through neural network-based decoding of the dequantized first feature data.

In one embodiment, the acquisition unit 510 applies the second feature data to the neural network (330; 350; 410) to obtain quantization data and probability data, and modifies the probability data based on sample values of the quantization data. It may include an AI control unit 650.

In one embodiment, the acquisition unit 510 may include an entropy decoding unit 610 that obtains quantized first feature data by applying entropy decoding based on modified probability data to bits included in the bitstream. You can.

In one embodiment, the acquisition unit 510 may include an inverse quantization unit 630 that inversely quantizes the quantized first feature data according to sample values of the quantization data to obtain the inverse quantized first feature data. .

The image encoding apparatus 1400 according to an embodiment may include a prediction encoder 1410 that acquires first feature data through neural network-based encoding of the current image 100.

In one embodiment, the video encoding device 1400 may include a generator 1420 that generates a bitstream corresponding to the current video 100.

In one embodiment, the generator 1420 applies the first feature data to the first neural network 310 to obtain second feature data for the first feature data, and applies the second feature data to the second neural network 330. ;350;410) to obtain quantization data and probability data, and may include an AI control unit 1510 that modifies the probability data based on sample values of the quantization data.

In one embodiment, the generator 1420 may include a quantization unit 1530 that quantizes the first feature data according to sample values of the quantization data to obtain quantized first feature data.

In one embodiment, the generator 1420 applies entropy encoding based on modified probability data to the quantized first feature data to generate an entropy code that generates a bitstream including bits corresponding to the quantized first feature data. It may include an encoder 1550.

In one embodiment, the bitstream may include bits corresponding to second characteristic data.

The image encoding device 1400, the image decoding device 500, and a method using them according to an embodiment can efficiently quantize/dequantize feature data generated through AI-based encoding of an image.

The video encoding device 1400, the video decoding device 500, and the method using them according to an embodiment can reduce the bit rate of the bitstream generated as a result of video encoding and improve the quality of the reconstructed video. .

The image encoding device 1400, the image decoding device 500, and the method using them according to an embodiment can improve the efficiency of entropy coding.

The video encoding device 1400, the video decoding device 500, and a method using them according to an embodiment can provide an AI-based end-to-end encoding/decoding system.

Claims (15)

Obtaining second feature data for first feature data obtained through neural network-based encoding of the current image 100 from a bitstream (S1310);
Obtaining quantization data and probability data by applying the second feature data to a neural network (330; 350; 410) (S1320);
Modifying the probability data based on sample values of the quantization data (S1330);
Obtaining quantized first feature data by applying entropy decoding based on the modified probability data to the bits included in the bitstream (S1340);
Obtaining dequantized first feature data by dequantizing the quantized first feature data according to sample values of the quantization data (S1350); and
A method of decoding an image, comprising restoring the current image through neural network-based decoding of the dequantized first feature data (S1360).
According to paragraph 1,
The quantized data is,
A method of decoding an image, including a quantization parameter or a quantization step size as a sample.
According to paragraph 1,
Sample values of the modified probability data are,
A method of decoding an image, indicating the probability of a value that a sample of the quantized first feature data may have.
According to paragraph 3,
Sample values of the modified probability data are,
A method for decoding an image, indicating an average and standard deviation corresponding to a sample of the quantized first feature data.
According to paragraph 4,
A method of decoding an image in which the probability of a value that a sample of the quantized first feature data may have is derived by applying the average and standard deviation indicated by the sample value of the modified probability data to a predetermined probability model.
According to paragraph 1,
The step of modifying the probability data is,
A method for decoding an image, comprising dividing sample values of the probability data into sample values of the quantization data.
According to paragraph 1,
The first characteristic data is,
Feature data (k) obtained by applying the current image 100 to the video encoder 12, feature data obtained by applying the current image 100 and the previous reconstructed image 200 to the optical flow encoder 42 (w) or feature data (v) obtained by applying the residual image (r) corresponding to the current image (100) to the residual encoder (52).
According to paragraph 1,
The neural network (330;350;410) is,
A method for decoding an image, trained based on at least one of the bitrate of the bitstream or the difference between a current training image and a current reconstructed training image.
According to paragraph 1,
By applying the second feature data to the neural network (330; 410), a plurality of probability data and a plurality of weights are obtained,
The plurality of probability data are modified based on sample values of the quantization data,
A method for decoding an image, wherein the probability of a value that a sample of the quantized first feature data can have is determined by combining the plurality of modified probability data according to the plurality of weights.
A computer-readable recording medium on which a program for performing the method of any one of claims 1 to 9 is recorded.
Applying first feature data obtained through neural network-based encoding of the current image 100 to the first neural network 310 to obtain second feature data for the first feature data (S1710);
Obtaining quantization data and probability data by applying the second feature data to a second neural network (330; 350; 410) (S1720);
Modifying the probability data based on sample values of the quantization data (S1730);
Obtaining quantized first feature data by quantizing the first feature data according to sample values of the quantization data (S1740); and
Generating a bitstream including bits corresponding to the quantized first feature data by applying entropy encoding based on the modified probability data to the quantized first feature data (S1750),
The bitstream includes bits corresponding to the second feature data.
According to clause 11,
The second neural network (330;350;410) is,
a quantization neural network 350 that outputs the quantization data from the second feature data; and
An image encoding method comprising a probabilistic neural network 330 that outputs the probability data from the second feature data.
According to clause 11,
The step of modifying the probability data is,
An image encoding method comprising dividing sample values of the probability data into sample values of the quantization data.
An acquisition unit 510 that acquires second feature data for first feature data obtained through neural network-based encoding of the current image 100 from a bitstream; and
A prediction decoder 530 that restores the current image 100 through neural network-based decoding of the dequantized first feature data,
The acquisition unit 510,
an AI control unit 650 that applies the second feature data to a neural network (330; 350; 410) to obtain quantization data and probability data, and modifies the probability data based on sample values of the quantization data;
an entropy decoding unit 610 that obtains quantized first feature data by applying entropy decoding based on the modified probability data to the bits included in the bitstream; and
An image decoding device comprising an inverse quantization unit 630 that inversely quantizes the quantized first feature data according to sample values of the quantized data to obtain the inverse quantized first feature data.
a prediction encoding unit 1410 that acquires first feature data through neural network-based encoding of the current image 100; and
It includes a generator 1420 that generates a bitstream corresponding to the current image 100,
The generating unit 1420,
Applying the first feature data to the first neural network 310 to obtain second feature data for the first feature data, and applying the second feature data to the second neural network (330; 350; 410) to quantize An AI control unit 1510 that acquires data and probability data and modifies the probability data based on sample values of the quantization data;
a quantization unit 1530 that quantizes the first feature data according to sample values of the quantization data to obtain quantized first feature data; and
and an entropy encoding unit 1550 that generates the bitstream including bits corresponding to the quantized first feature data by applying entropy coding based on the modified probability data to the quantized first feature data. ,
The bitstream includes bits corresponding to the second feature data.

Images