JP2019201255A - Image filter device, image decoding device, and image encoding device - Google Patents

Abstract

To convert a quantization parameter into a parameter suitable for filtering using a neural network.SOLUTION: The image filter device includes: a neural network unit (404), to which one or more first type of input image data is input in CNN filters (107, 305), for outputting one or more first-type output image data with luminance or color difference used as a pixel value; a conversion unit (402), to which one or more second type of input image data with a quantization parameter as a pixel value is input, for nonlinearly converting the second type of input image data to generate a second type of input image data after the conversion; The second type of input image data after the conversion is also input to the neural network unit.SELECTED DRAWING: Figure 5

Description

  Embodiments described herein relate generally to an image filter device, an image decoding device, and an image encoding device.

  In order to efficiently transmit or record a moving image, a moving image encoding device that generates encoded data by encoding the moving image, and a moving image that generates a decoded image by decoding the encoded data An image decoding device is used.

  Specific examples of the moving image encoding method include a method proposed in H.264 / AVC and HEVC (High-Efficiency Video Coding).

  In such a moving image coding system, an image (picture) constituting a moving image is a slice obtained by dividing the image, a coding tree unit (CTU: Coding Tree Unit) obtained by dividing the slice. ), A coding unit obtained by dividing the coding tree unit (sometimes referred to as a coding unit (CU)), and a prediction unit that is a block obtained by dividing the coding unit (PU) and a hierarchical structure composed of conversion units (TU), and encoded / decoded for each CU.

  In such a moving image coding method, a predicted image is usually generated based on a local decoded image obtained by encoding / decoding an input image, and the predicted image is generated from the input image (original image). A prediction residual obtained by subtraction (sometimes referred to as “difference image” or “residual image”) is encoded. Examples of the method for generating a predicted image include inter-screen prediction (inter prediction) and intra-screen prediction (intra prediction).

  Further, Non-Patent Document 1 can be cited as a technique for encoding and decoding moving images in recent years.

  Non-patent document 2 is cited as a technique using a neural network called Variable-filter-size Residue-learning CNN (VRCNN).

"Algorithm Description of Joint Exploration Test Model 5", JVET-E1001, Joint Video Exploration Team (JVET) of ITU-T SG 16 WP 3 and ISO / IEC JTC 1 / SC 29 / WG 11, 12-20 January 2017 "A Convolutional Neural Network Approach for Post-Processing in HEVC Intra Coding"

  However, the filtering process using the above-described neural network is not suitable for the quantization parameter.

  One embodiment of the present invention has been made in view of the above problem, and an object thereof is to convert a quantization parameter into a parameter suitable for filter processing using a neural network.

  In order to solve the above-described problem, an image filter device according to one embodiment of the present invention receives at least one or more first-type input image data having a pixel value of luminance or color difference, and sets the luminance or color difference to pixel A neural network unit that outputs one or a plurality of first type output image data as a value, and one or a plurality of second type input image data that uses a quantization parameter as a pixel value are input, and the second type A conversion unit that generates a second type of input image data after conversion by nonlinearly converting the input image data, and the neural network unit also receives the second type of input image data after conversion. Is done.

  According to one embodiment of the present invention, the quantization parameter can be used for filter processing using a neural network.

It is a figure which shows the hierarchical structure of the data of the encoding stream which concerns on 1st Embodiment etc. It is a figure which shows the pattern of PU division | segmentation mode. (A) to (h) show the partition shapes when the PU partitioning modes are 2Nx2N, 2NxN, 2NxnU, 2NxnD, Nx2N, nLx2N, nRx2N, and NxN, respectively. It is a block diagram which shows the structure of the image coding apparatus which concerns on 1st Embodiment etc. It is a block diagram which shows the structure of the image decoding apparatus which concerns on 1st Embodiment etc. It is the schematic which shows the structure of the image filter apparatus concerning 1st Embodiment etc., and the input / output of image data. It is the graph which showed the relationship between the conversion quantization parameter (qPNL) which the conversion part concerning 1st Embodiment etc. nonlinearly converted, and the quantization parameter (qP). (A) is a graph when using a conversion formula for decimal point arithmetic, and (b) is a graph when using a conversion formula for integer arithmetic. It is the graph which showed the relationship between the conversion quantization parameter (qPNL) which the conversion part concerning 1st Embodiment etc. nonlinearly converted, and the quantization parameter (qP). (A) is a graph when using a conversion formula for decimal point arithmetic, and (b) is a graph when using a conversion formula for integer arithmetic. It is the graph which showed the relationship between the conversion quantization parameter (qPNL) which the conversion part concerning 1st Embodiment etc. nonlinearly converted, and the quantization parameter (qP). (A) is a graph when using a logarithmic conversion equation, (b) is a graph when using an exponential conversion equation, and (c) is a graph when using an inverse conversion equation. . It is the schematic which shows the structure of the image filter apparatus concerning 1st Embodiment etc., and the input / output of image data. It is a figure for demonstrating an example of a quantization parameter. It is a conceptual diagram which shows an example of the image filter apparatus concerning 4th Embodiment etc. It is a conceptual diagram which shows the modification of an example of the image filter apparatus which concerns on 4th Embodiment etc. It is a conceptual diagram which shows an example of the image filter apparatus concerning 5th Embodiment etc. It is a conceptual diagram which shows the modification of the image filter apparatus concerning 6th Embodiment etc. It is the figure shown about the structure of the transmitter which mounts the image coding apparatus which concerns on each embodiment, and the receiver which mounts an image decoding apparatus. (A) shows a transmission device equipped with an image encoding device, and (b) shows a reception device equipped with an image decoding device. It is the figure shown about the structure of the recording device carrying the image coding apparatus which concerns on each embodiment, and the reproducing | regenerating apparatus carrying an image decoding apparatus. (A) shows a recording device equipped with an image encoding device, and (b) shows a playback device equipped with an image decoding device. It is the schematic which shows the structure of the image transmission system which concerns on each embodiment. It is a conceptual diagram which shows the other example of the input / output of the image filter apparatus which concerns on 1st Embodiment etc. It is a figure for demonstrating an example by which a quantization parameter is corrected into a correction quantization parameter. (A) shows the target block before correction, and (b) shows the target block after correction. It is a figure for demonstrating an example by which a quantization parameter is corrected into a correction quantization parameter. (A) shows the target block before correction, and (b) shows the target block after correction. It is a figure for demonstrating an example by which a quantization parameter is corrected into a correction quantization parameter. (A) shows the target block before correction, and (b) shows the target block after correction. It is a figure for demonstrating an example by which a quantization parameter is corrected into a correction quantization parameter. (A) shows the target block before correction, and (b) shows the target block after correction. It is a figure for demonstrating an example of discrimination | determination of the block boundary of the adjacent blocks P and Q. FIG. (A) shows an example of discriminating a horizontal boundary, and (b) shows an example of discriminating a vertical boundary.

<First Embodiment>
Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 17 is a schematic diagram illustrating a configuration of the image transmission system 1 according to the present embodiment.

  The image transmission system 1 is a system that transmits a code obtained by encoding an image to be encoded, decodes the transmitted code, and displays an image. The image transmission system 1 includes an image encoding device (moving image encoding device) 11, a network 21, an image decoding device (moving image decoding device) 31, and an image display device 41.

  An image T indicating a single layer image or a plurality of layers of images is input to the image encoding device 11. A layer is a concept used to distinguish a plurality of pictures when there are one or more pictures constituting a certain time. For example, when the same picture is encoded with a plurality of layers having different image quality and resolution, scalable encoding is performed, and when a picture of a different viewpoint is encoded with a plurality of layers, view scalable encoding is performed. When prediction is performed between pictures of a plurality of layers (inter-layer prediction, inter-view prediction), encoding efficiency is greatly improved. Further, even when prediction is not performed (simultaneous casting), encoded data can be collected.

  The network 21 transmits the encoded stream Te generated by the image encoding device 11 to the image decoding device 31. The network 21 is the Internet, a wide area network (WAN), a small network (LAN), or a combination thereof. The network 21 is not necessarily limited to a bidirectional communication network, and may be a unidirectional communication network that transmits broadcast waves such as terrestrial digital broadcasting and satellite broadcasting. The network 21 may be replaced with a storage medium that records an encoded stream Te such as a DVD (Digital Versatile Disc) or a BD (Blue-ray Disc).

  The image decoding device 31 decodes each encoded stream Te transmitted by the network 21, and generates one or a plurality of decoded images Td decoded.

  The image display device 41 displays all or part of one or a plurality of decoded images Td generated by the image decoding device 31. The image display device 41 includes, for example, a display device such as a liquid crystal display or an organic EL (Electro-luminescence) display. In addition, in the spatial scalable coding and SNR scalable coding, when the image decoding device 31 and the image display device 41 have a high processing capability, an enhancement layer image with a high image quality is displayed and only a lower processing capability is provided. Displays a base layer image that does not require higher processing capability and display capability as an extension layer.

<Operator>
The operators used in this specification are described below.

  >> is right bit shift, << is left bit shift, & is bitwise AND, | is bitwise OR, | = is OR assignment operator).

  x? y: z is a ternary operator that takes y when x is true (other than 0) and takes z when x is false (0).

  Clip3 (a, b, c) is a function that clips c to a value greater than or equal to a and less than or equal to b, returns a if c <a, returns b if c> b, otherwise Is a function that returns c (where a <= b).

<Structure of encoded stream Te>
Prior to detailed description of the image encoding device 11 and the image decoding device 31 according to the present embodiment, a data structure of an encoded stream Te generated by the image encoding device 11 and decoded by the image decoding device 31 will be described. .

  FIG. 1 is a diagram illustrating a hierarchical structure of data in the encoded stream Te. The encoded stream Te illustratively includes a sequence and a plurality of pictures constituting the sequence. (A) to (f) of FIG. 1 respectively show an encoded video sequence defining a sequence SEQ, an encoded picture defining a picture PICT, an encoded slice defining a slice S, and an encoded slice defining a slice data It is a figure which shows the coding unit (Coding Unit; CU) contained in the coding tree unit contained in data and coding slice data, and a coding tree unit.

(Encoded video sequence)
In the encoded video sequence, a set of data referred to by the image decoding device 31 for decoding the sequence SEQ to be processed is defined. As shown in FIG. 1A, the sequence SEQ includes a video parameter set (Video Parameter Set), a sequence parameter set SPS (Sequence Parameter Set), a picture parameter set PPS (Picture Parameter Set), a picture PICT, and an addition. Includes SEI (Supplemental Enhancement Information). Here, the value indicated after # indicates the layer ID. Although FIG. 1 shows an example in which encoded data of # 0 and # 1, that is, layer 0 and layer 1, exists, the type of layer and the number of layers are not dependent on this.

  The video parameter set VPS is a set of encoding parameters common to a plurality of moving images, a plurality of layers included in the moving image, and encoding parameters related to individual layers in a moving image composed of a plurality of layers. A set is defined.

  In the sequence parameter set SPS, a set of encoding parameters referred to by the image decoding device 31 in order to decode the target sequence is defined. For example, the width and height of the picture are defined. A plurality of SPSs may exist. In that case, one of a plurality of SPSs is selected from the PPS.

  In the picture parameter set PPS, a set of encoding parameters referred to by the image decoding device 31 in order to decode each picture in the target sequence is defined. For example, a quantization width reference value (pic_init_qp_minus26) used for picture decoding and a flag (weighted_pred_flag) indicating application of weighted prediction are included. There may be a plurality of PPSs. In that case, one of a plurality of PPSs is selected from each picture in the target sequence.

(Encoded picture)
In the coded picture, a set of data referred to by the image decoding device 31 in order to decode the picture PICT to be processed is defined. As shown in FIG. 1B, the picture PICT includes slices S0 to S _NS-1 (NS is the total number of slices included in the picture PICT).

Hereinafter, when it is not necessary to distinguish each of the slices S0 to SNS _-1 , the subscripts may be omitted. The same applies to data included in an encoded stream Te described below and to which other subscripts are attached.

(Encoded slice)
In the coded slice, a set of data referred to by the image decoding device 31 for decoding the slice S to be processed is defined. As shown in FIG. 1C, the slice S includes a slice header SH and slice data SDATA.

  The slice header SH includes an encoding parameter group that is referred to by the image decoding device 31 in order to determine a decoding method of the target slice. Slice type designation information (slice_type) for designating a slice type is an example of an encoding parameter included in the slice header SH.

  As slice types that can be specified by the slice type specification information, (1) I slice using only intra prediction at the time of encoding, (2) P slice using unidirectional prediction or intra prediction at the time of encoding, (3) B-slice using unidirectional prediction, bidirectional prediction, or intra prediction at the time of encoding may be used.

  Note that the slice header SH may include a reference (pic_parameter_set_id) to the picture parameter set PPS included in the encoded video sequence.

(Encoded slice data)
In the encoded slice data, a set of data referred to by the image decoding device 31 for decoding the slice data SDATA to be processed is defined. The slice data SDATA includes a coding tree unit (CTU) as shown in FIG. A CTU is a block of a fixed size (for example, 64x64) that constitutes a slice, and is sometimes called a maximum coding unit (LCU: Large Coding Unit).

(Encoding tree unit)
As shown in (e) of FIG. 1, a set of data referred to by the image decoding device 31 in order to decode the encoding tree unit to be processed is defined. The coding tree unit is divided by recursive quadtree division. A tree-structured node obtained by recursive quadtree partitioning is referred to as a coding node (CN). An intermediate node of the quadtree is an encoding node, and the encoding tree unit itself is defined as the highest encoding node. The CTU includes a split flag (cu_split_flag), and when cu_split_flag is 1, it is split into four coding nodes CN. When cu_split_flag is 0, the coding node CN is not divided and has one coding unit (CU: Coding Unit) as a node. The encoding unit CU is a terminal node of the encoding node and is not further divided. The encoding unit CU is a basic unit of the encoding process.

  When the size of the coding tree unit CTU is 64x64 pixels, the size of the coding unit can be any of 64x64 pixels, 32x32 pixels, 16x16 pixels, and 8x8 pixels.

(Encoding unit)
As shown in (f) of FIG. 1, a set of data referred to by the image decoding device 31 in order to decode an encoding unit to be processed is defined. Specifically, the encoding unit includes a prediction tree, a conversion tree, and a CU header CUH. In the CU header, a prediction mode, a division method (PU division mode), and the like are defined.

  In the prediction tree, prediction information (a reference picture index, a motion vector, etc.) of each prediction unit (PU) obtained by dividing the coding unit into one or a plurality of parts is defined. In other words, the prediction unit is one or a plurality of non-overlapping areas constituting the encoding unit. The prediction tree includes one or a plurality of prediction units obtained by the above-described division. Hereinafter, a prediction unit obtained by further dividing the prediction unit is referred to as a “sub-block”. The sub block is composed of a plurality of pixels. When the sizes of the prediction unit and the sub-block are equal, the number of sub-blocks in the prediction unit is one. If the prediction unit is larger than the size of the sub-block, the prediction unit is divided into sub-blocks. For example, when the prediction unit is 8 × 8 and the sub-block is 4 × 4, the prediction unit is divided into four sub-blocks that are divided into two horizontally and two vertically.

  The prediction process may be performed for each prediction unit (sub-block).

  Broadly speaking, there are two types of division in the prediction tree: intra prediction and inter prediction. Intra prediction is prediction within the same picture, and inter prediction refers to prediction processing performed between different pictures (for example, between display times and between layer images).

  In the case of intra prediction, there are 2Nx2N (the same size as the coding unit) and NxN division methods.

Also, in the case of inter prediction, the division method is encoded by the PU division mode (part_mode) of the encoded data, 2Nx2N (the same size as the encoding unit), 2NxN, 2NxnU, 2NxnD, Nx2N, nLx2N, nRx2N, and NxN etc. 2NxN and Nx2N indicate 1: 1 symmetrical division,
2NxnU, 2NxnD and nLx2N, nRx2N show a 1: 3, 3: 1 asymmetric partitioning. The PUs included in the CU are expressed as PU0, PU1, PU2, and PU3 in this order.

  2A to 2H specifically illustrate the shape of the partition (the position of the boundary of the PU partition) in each PU partition mode. 2A shows a 2Nx2N partition, and FIGS. 2B, 2C, and 2D show 2NxN, 2NxnU, and 2NxnD partitions (horizontal partitions), respectively. (E), (f), and (g) show partitions (vertical partitions) in the case of Nx2N, nLx2N, and nRx2N, respectively, and (h) shows an NxN partition. The horizontal partition and the vertical partition are collectively referred to as a rectangular partition, and 2Nx2N and NxN are collectively referred to as a square partition.

  In the transform tree, the encoding unit is divided into one or a plurality of transform units, and the position and size of each transform unit are defined. In other words, a transform unit is one or more non-overlapping areas that make up a coding unit. The conversion tree includes one or a plurality of conversion units obtained by the above division.

  There are two types of division in the conversion tree: one in which an area having the same size as that of the encoding unit is allocated as a conversion unit, and the other in division by recursive quadtree division, similar to the above-described CU division.

  The conversion process is performed for each conversion unit.

(Prediction parameter)
A prediction image of a prediction unit (PU: Prediction Unit) is derived from a prediction parameter associated with the PU. The prediction parameters include a prediction parameter for intra prediction or a prediction parameter for inter prediction. Hereinafter, prediction parameters for inter prediction (inter prediction parameters) will be described. The inter prediction parameter includes prediction list use flags predFlagL0 and predFlagL1, reference picture indexes refIdxL0 and refIdxL1, and motion vectors mvL0 and mvL1. The prediction list use flags predFlagL0 and predFlagL1 are flags indicating whether or not reference picture lists called L0 list and L1 list are used, respectively, and a reference picture list corresponding to a value of 1 is used. In this specification, when “flag indicating whether or not it is XX” is described, when the flag is not 0 (for example, 1) is XX, 0 is not XX, and logical negation, logical product, etc. 1 is treated as true and 0 is treated as false (the same applies hereinafter). However, other values can be used as true values and false values in an actual apparatus or method.

  Syntax elements for deriving inter prediction parameters included in the encoded data include, for example, PU partition mode part_mode, merge flag merge_flag, merge index merge_idx, inter prediction identifier inter_pred_idc, reference picture index refIdxLX, prediction vector index mvp_LX_idx, There is a difference vector mvdLX.

(Reference picture list)
The reference picture list is a list including reference pictures stored in the reference picture memory 306.

(Merge prediction and AMVP prediction)
The prediction parameter decoding (encoding) method includes a merge prediction (merge) mode and an AMVP (Adaptive Motion Vector Prediction) mode. The merge flag merge_flag is a flag for identifying these. The merge prediction mode is a mode in which the prediction list use flag predFlagLX (or inter prediction identifier inter_pred_idc), the reference picture index refIdxLX, and the motion vector mvLX are not included in the encoded data and are derived from the prediction parameters of already processed neighboring PUs. The AMVP mode is a mode in which the inter prediction identifier inter_pred_idc, the reference picture index refIdxLX, and the motion vector mvLX are included in the encoded data. The motion vector mvLX is encoded as a prediction vector index mvp_LX_idx for identifying the prediction vector mvpLX and a difference vector mvdLX.

(Motion vector)
The motion vector mvLX indicates a shift amount between blocks on two different pictures. A prediction vector and a difference vector related to the motion vector mvLX are referred to as a prediction vector mvpLX and a difference vector mvdLX, respectively.

(Configuration of image decoding device)
Next, the configuration of the image decoding device 31 according to the present embodiment will be described. FIG. 4 is a schematic diagram illustrating a configuration of the image decoding device 31 according to the present embodiment. The image decoding device 31 includes an entropy decoding unit 301, a prediction parameter decoding unit (prediction image decoding device) 302, a CNN (Convolution
al Neural Network (convolutional neural network) filter 305, reference picture memory 306, prediction parameter memory 307, prediction image generation unit (prediction image generation device) 308, inverse quantization / inverse conversion unit 311, and addition unit 312. Is done.

  The prediction parameter decoding unit 302 includes an inter prediction parameter decoding unit 303 and an intra prediction parameter decoding unit 304. The predicted image generation unit 308 includes an inter predicted image generation unit 309 and an intra predicted image generation unit 310.

  The entropy decoding unit 301 performs entropy decoding on an encoded stream Te input from the outside, and separates and decodes individual codes (syntax elements). The separated codes include prediction information for generating a prediction image and residual information for generating a difference image.

  The entropy decoding unit 301 outputs a part of the separated code to the prediction parameter decoding unit 302. Some of the separated codes are, for example, a prediction mode predMode, a PU partition mode part_mode, a merge flag merge_flag, a merge index merge_idx, an inter prediction identifier inter_pred_idc, a reference picture index refIdxLX, a prediction vector index mvp_LX_idx, and a difference vector mvdLX. Control of which code is decoded is performed based on an instruction from the prediction parameter decoding unit 302. The entropy decoding unit 301 outputs the quantized coefficient to the inverse quantization / inverse transform unit 311. In the coding process, this quantization coefficient is applied to the residual signal by DCT (Discrete Cosine Transform), DST (Discrete Sine Transform), KLT (Karyhnen Loeve Transform) It is a coefficient obtained by performing frequency conversion such as

  Based on the code input from the entropy decoding unit 301, the inter prediction parameter decoding unit 303 refers to the prediction parameter stored in the prediction parameter memory 307 and decodes the inter prediction parameter.

  The inter prediction parameter decoding unit 303 outputs the decoded inter prediction parameter to the prediction image generation unit 308 and stores it in the prediction parameter memory 307. Details of the inter prediction parameter decoding unit 303 will be described later.

  Based on the code input from the entropy decoding unit 301, the intra prediction parameter decoding unit 304 refers to the prediction parameter stored in the prediction parameter memory 307 and decodes the intra prediction parameter. The intra prediction parameter is a parameter used in a process of predicting a CU within one picture, for example, an intra prediction mode IntraPredMode. The intra prediction parameter decoding unit 304 outputs the decoded intra prediction parameter to the prediction image generation unit 308 and stores it in the prediction parameter memory 307.

  The intra prediction parameter decoding unit 304 may derive different intra prediction modes depending on the luminance and color difference. In this case, the intra prediction parameter decoding unit 304 decodes the luminance prediction mode IntraPredModeY as the luminance prediction parameter and the color difference prediction mode IntraPredModeC as the color difference prediction parameter. The luminance prediction mode IntraPredModeY is a 35 mode, and corresponds to planar prediction (0), DC prediction (1), and direction prediction (2 to 34). The color difference prediction mode IntraPredModeC uses any one of the planar prediction (0), the DC prediction (1), the direction prediction (2 to 34), and the LM mode (35). The intra prediction parameter decoding unit 304 decodes a flag indicating whether IntraPredModeC is the same mode as the luminance mode. If the flag indicates that the mode is the same as the luminance mode, IntraPredModeC is assigned to IntraPredModeC, and the flag indicates luminance. If the mode is different from the mode, planar prediction (0), DC prediction (1), direction prediction (2 to 34), and LM mode (35) may be decoded as IntraPredModeC.

The CNN filter 305 acquires the quantization parameter and the prediction parameter from the entropy decoding unit 301, uses the decoded image of the CU generated by the addition unit 312 as an input image (pre-filter image), performs processing on the pre-filter image, and outputs Output image (filtered image). The CNN filter 305 has the same function as the CNN filter 107 included in the image encoding device 11.

  The reference picture memory 306 stores the decoded image of the CU generated by the adding unit 312 at a predetermined position for each decoding target picture and CU.

  The prediction parameter memory 307 stores the prediction parameter at a predetermined position for each decoding target picture and prediction unit (or sub-block, fixed-size block, pixel). Specifically, the prediction parameter memory 307 stores the inter prediction parameter decoded by the inter prediction parameter decoding unit 303, the intra prediction parameter decoded by the intra prediction parameter decoding unit 304, and the prediction mode predMode separated by the entropy decoding unit 301. . The stored inter prediction parameters include, for example, a prediction list utilization flag predFlagLX (inter prediction identifier inter_pred_idc), a reference picture index refIdxLX, and a motion vector mvLX.

  The prediction image generation unit 308 receives the prediction mode predMode input from the entropy decoding unit 301, and also receives prediction parameters from the prediction parameter decoding unit 302. Further, the predicted image generation unit 308 reads a reference picture from the reference picture memory 306. The prediction image generation unit 308 generates a prediction image of a PU or sub-block using the input prediction parameter and the read reference picture (reference picture block) in the prediction mode indicated by the prediction mode predMode.

  Here, when the prediction mode predMode indicates the inter prediction mode, the inter prediction image generation unit 309 uses the inter prediction parameter input from the inter prediction parameter decoding unit 303 and the read reference picture (reference picture block). To generate a prediction image of a PU or sub-block.

  For the reference picture list (L0 list or L1 list) for which the prediction list use flag predFlagLX is 1, the inter predicted image generation unit 309 uses the decoding target PU as a reference from the reference picture indicated by the reference picture index refIdxLX. The reference picture block at the position indicated by mvLX is read from the reference picture memory 306. The inter prediction image generation unit 309 performs prediction based on the read reference picture block to generate a prediction image of the PU. The inter prediction image generation unit 309 outputs the generated prediction image of the PU to the addition unit 312. Here, a reference picture block is a set of pixels on a reference picture (usually called a block because it is a rectangle), and is an area that is referred to in order to generate a predicted image of a PU or sub-block.

  When the prediction mode predMode indicates the intra prediction mode, the intra predicted image generation unit 310 performs intra prediction using the intra prediction parameter input from the intra prediction parameter decoding unit 304 and the read reference picture. Specifically, the intra predicted image generation unit 310 reads, from the reference picture memory 306, neighboring PUs that are pictures to be decoded and are in a predetermined range from the decoding target PUs among the PUs that have already been decoded. The predetermined range is, for example, one of the left, upper left, upper, and upper right adjacent PUs when the decoding target PU sequentially moves in the so-called raster scan order, and differs depending on the intra prediction mode. The raster scan order is an order in which each row is sequentially moved from the left end to the right end in each picture from the upper end to the lower end.

  The intra predicted image generation unit 310 performs prediction in the prediction mode indicated by the intra prediction mode IntraPredMode based on the read adjacent PU, and generates a predicted image of the PU. The intra predicted image generation unit 310 outputs the generated predicted image of the PU to the adding unit 312.

  When the intra prediction parameter decoding unit 304 derives an intra prediction mode different in luminance and color difference, the intra prediction image generation unit 310 performs planar prediction (0), DC prediction (1), direction according to the luminance prediction mode IntraPredModeY. The prediction image of the luminance PU is generated by any one of the predictions (2 to 34), and the planar prediction (0), the DC prediction (1), the direction prediction (2 to 34), and the LM mode according to the color difference prediction mode IntraPredModeC. A predicted image of the color difference PU is generated by any of (35).

  The inverse quantization / inverse transform unit 311 performs inverse quantization on the quantization coefficient input from the entropy decoding unit 301 to obtain a transform coefficient. The inverse quantization / inverse transform unit 311 performs inverse frequency transform such as inverse DCT, inverse DST, and inverse KLT on the obtained transform coefficient, and calculates a residual signal. The inverse quantization / inverse transform unit 311 outputs the calculated residual signal to the adder 312.

  The inverse quantization / inverse transform unit 311 calculates the quantization step obtained by QP = 2 ^ (qP / 6) from the quantization parameter (qP) decoded with reference to the encoded stream Te, The quantized coefficient input from the entropy decoding unit 301 is inversely quantized using the calculated quantization step (QP).

  The addition unit 312 adds the prediction image of the PU input from the inter prediction image generation unit 309 or the intra prediction image generation unit 310 and the residual signal input from the inverse quantization / inverse conversion unit 311 for each pixel, Generate a decoded PU image. The adding unit 312 stores the generated decoded image of the PU in the reference picture memory 306, and outputs a decoded image Td obtained by integrating the generated decoded image of the PU for each picture to the outside.

(Configuration of image encoding device)
Next, the configuration of the image encoding device 11 according to the present embodiment will be described. FIG. 3 is a block diagram illustrating a configuration of the image encoding device 11 according to the present embodiment. The image encoding device 11 includes a predicted image generation unit 101, a subtraction unit 102, a transform / quantization unit 103, an entropy encoding unit 104, an inverse quantization / inverse transform unit 105, an addition unit 106, a CNN (Convolutional Neural Network, convolution). Neural network) filter 107, prediction parameter memory (prediction parameter storage unit, frame memory) 108, reference picture memory (reference image storage unit, frame memory) 109, encoding parameter determination unit 110, and prediction parameter encoding unit 111. Composed. The prediction parameter encoding unit 111 includes an inter prediction parameter encoding unit 112 and an intra prediction parameter encoding unit 113.

  For each picture of the image T, the predicted image generation unit 101 generates a predicted image P of the prediction unit PU for each encoding unit CU that is an area obtained by dividing the picture. Here, the predicted image generation unit 101 reads a decoded block from the reference picture memory 109 based on the prediction parameter input from the prediction parameter encoding unit 111. The prediction parameter input from the prediction parameter encoding unit 111 is, for example, a motion vector in the case of inter prediction. The predicted image generation unit 101 reads a block at a position on the reference image indicated by the motion vector with the target PU as a starting point. In the case of intra prediction, the prediction parameter is, for example, an intra prediction mode. A pixel value of an adjacent PU used in the intra prediction mode is read from the reference picture memory 109, and a predicted image P of the PU is generated. The predicted image generation unit 101 generates a predicted image P of the PU using one prediction method among a plurality of prediction methods for the read reference picture block. The predicted image generation unit 101 outputs the generated predicted image P of the PU to the subtraction unit 102.

  Note that since the predicted image generation unit 101 has the same operation as the predicted image generation unit 308 already described, description thereof is omitted here.

  The predicted image generation unit 101 generates a predicted image P of the PU based on the pixel value of the reference block read from the reference picture memory, using the parameter input from the prediction parameter encoding unit. The predicted image generated by the predicted image generation unit 101 is output to the subtraction unit 102 and the addition unit 106.

  The subtraction unit 102 subtracts the signal value of the predicted image P of the PU input from the predicted image generation unit 101 from the pixel value of the corresponding PU of the image T, and generates a residual signal. The subtraction unit 102 outputs the generated residual signal to the transform / quantization unit 103.

  The transformation / quantization unit 103 performs frequency transformation on the residual signal input from the subtraction unit 102 and calculates a transformation coefficient. The transform / quantization unit 103 quantizes the calculated transform coefficient to obtain a quantized coefficient. The transform / quantization unit 103 outputs the obtained quantization coefficient to the entropy coding unit 104 and the inverse quantization / inverse transform unit 105.

  The entropy encoding unit 104 receives the quantization coefficient from the transform / quantization unit 103 and the encoding parameter from the prediction parameter encoding unit 111. The input coding parameters include, for example, codes such as quantization parameters, depth information (partition information), reference picture index refIdxLX, prediction vector index mvp_LX_idx, difference vector mvdLX, prediction mode predMode, and merge index merge_idx.

  The entropy encoding unit 104 generates an encoded stream Te by entropy encoding the input quantization coefficient and encoding parameter, and outputs the generated encoded stream Te to the outside.

  The inverse quantization / inverse transform unit 105 inversely quantizes the quantization coefficient input from the transform / quantization unit 103 to obtain a transform coefficient. The inverse quantization / inverse transform unit 105 performs inverse frequency transform on the obtained transform coefficient to calculate a residual signal. The inverse quantization / inverse transform unit 105 outputs the calculated residual signal to the addition unit 106.

  The addition unit 106 adds the signal value of the prediction image P of the PU input from the prediction image generation unit 101 and the signal value of the residual signal input from the inverse quantization / inverse conversion unit 105 for each pixel, and performs decoding. Generate an image. The adding unit 106 stores the generated decoded image in the reference picture memory 109.

(Configuration of image filter device)
The CNN filter 305 is an example of an image filter device according to the present embodiment. FIG. 5 shows the configuration of the image filter device. The image filter device includes a qP image generation unit 400, a neural network unit 404, and a conversion unit 402.

  Here, the CNN filter 305 (neural network unit 404) is a convolution layer (a neuron has a spatial position and is connected only to an input close to the spatial position, and is used in the product-sum operation of the neuron. A generic term for a neural network having at least a coefficient / kernel and a layer whose bias / offset does not depend on the position in the picture. In addition to the convolution layer, the CNN filter 305 is called a full connection layer (FCN). A neuron does not have a spatial position and is connected to all inputs, or a neuron called an LCN (Locally Connected Networks) layer. Is a configuration that has a spatial position and is connected only to an input close to the spatial position, but can include a configuration that does not share a weighting factor or a bias unlike CNN. In the CNN filter 305, the input size to the convolution layer and the output size from the convolution layer may be different. That is, the CNN filter 305 can include a convolution layer in which the output size is smaller than the input size by making the movement amount (step size) when moving the position to which the convolution filter is applied larger than 1. . In addition, a deconvolution layer (Deconvolution) in which the output size is larger than the input size can be included. The deconvolution layer is sometimes called a transposed convolution layer. The CNN filter 305 may include a pooling layer (Pooling), a dropout layer (DropOut) layer, and the like. The pooling layer is a layer that divides a large image into small windows, and obtains a representative value such as a maximum value or an average value according to each divided window, and the dropout layer has a fixed output (for example, output depending on the probability) 0) is a layer for adding randomness.

  FIG. 18 is a conceptual diagram showing another example of input / output of the CNN filter 305. In FIG. 18A, the pre-filter image has a luminance (Y) and quantization parameter (qP) channel, a first color difference (Cb) and quantization parameter (qP) channel, and a second color difference. (Cr) and quantization parameter (qP) channels are input to the CNN filter 305. The CNN filter 305 performs processing on the luminance (Y) and quantization parameter (qP) channels and outputs (Y ′), and outputs the first color difference (Cb) and quantization parameter (qP). The CNN filter 305-2 that processes (Cb ′) and outputs the (Cb ′), and the CNN filter 305 that processes the second color difference (Cr) and the quantization parameter (qP) and outputs (Cr ′). -3.

  In FIG. 18B, the pre-filter image includes the luminance (Y) and quantization parameter (qP) channels, the first color difference (Cb), the second color difference (Cr), and the quantization parameter. It is divided into (qP) channels and input to the CNN filter 305. The CNN filter 305 performs processing on the luminance (Y) and quantization parameter (qP) channels and outputs (Y ′), and outputs a first color difference (Cb) and a second color difference (Cr ) And a CNN filter 305-5 that processes the channel of the quantization parameter (qP) and outputs (Cb ′, Cr ′).

  Although not shown, the channels of luminance (Y), first color difference (Cb), second color difference (Cr), and quantization parameter (qP) are input to the CNN filter 305, and (Y ′, Cb ′, Cr CNN filter 305 that outputs') may be used. Also, four luminance (Y0, Y1, Y2, Y3), first color difference (Cb), second color difference (Cr), and quantization parameter (qP) channels are input to the CNN filter 305, and (Y0 CNN filter 305 that outputs ', Y1', Y2 ', Y3', Cb ', Cr') may be used. Of course, an RGB image (R, G, B) may also be used.

  Here, the reference parameter (encoding parameter) is not limited to the quantization parameter (qP), and one or more encoding parameters can be used. The CNN filter 305 includes a CNN filter 305-1, a CNN filter 305-2, and CNN filters 305-3, 305-4, and 305-5, which are configured using different means (circuits and software). It is not limited. For example, the CNN filter 305 may be configured by one unit among a plurality of different units (circuits and software), and may operate in different modes.

  Here, each of the CNN filters 305-1 to 305-3 includes, for example, a qP image generation unit, a conversion unit, and a neural network unit as described with reference to FIG.

  Here, each of the CNN filters 305-4 to 305-5 includes, for example, a qP image generation unit, a conversion unit, and a neural network unit as described with reference to FIG.

  Further, the reference parameter (encoding parameter) is not limited to the quantization parameter (qP), and one or more encoding parameters can be used.

  In the following embodiments, the processing may be performed with the channel configuration as described above. Further, the reference parameter (encoding parameter) is not limited to the quantization parameter (qP), and the CNN filter d1 can use one or more encoding parameters.

  As illustrated in FIG. 5, the qP image generation unit 400 refers to the quantization parameter (qP) specified for each block, and assigns the value of qP in the block as a pixel value to each pixel included in the block. Thus, a qP image (second type of input image data) is generated, and the generated qP image is supplied to the conversion unit 402. A description of the quantization parameter (qP) and a specific example of the qP image will be described later.

  The conversion unit 402 generates a converted qPNL image by nonlinearly converting each quantization parameter (qP) included in the qP image acquired from the qP image generation unit 400. The qPNL image after the nonlinear transformation is an image in which each pixel value is a quantization parameter (qP) after the nonlinear transformation.

  The nonlinear-transformed qPNL image (second-type input image data after conversion) generated by the conversion unit 402 is input to the neural network unit 404.

  The neural network unit 404 receives the decoded image (first type of input image data) and the image subjected to nonlinear transformation by the conversion unit 402, and outputs an output image (filtered image). The decoded image may be a one-channel image (eg, luminance component image {Y}, each color difference component image {Cb}, {Cr}), or a plurality of channel images (eg, luminance and color difference components). Image {Y, Cb}, {Y, Cr}, two color difference component images {Cb, Cr}, and luminance and two color difference images {Y, Cb, Cr}).

  The decoded image may be the output of the addition unit, or may be a decoded image after application of one or more filters (bilateral filter, deblocking filter, SAO, ALF, NLMF, etc.) after addition.

  In FIG. 5A, the qP image generation unit 400 generates qP two-dimensional data, and converts the two-dimensional data generated by the conversion unit 402. As shown in FIG. Alternatively, after converting qP in scalar units, it may be converted into qPNL two-dimensional data after conversion. In this configuration, the qP derivation unit 401 refers to the quantization parameter (qP) specified for each block and inputs the quantization parameter (qP) to the conversion unit 402. The conversion unit 402 converts a scalar qP value into a qPNL value. Further, the qP image generation unit 403 generates two-dimensional data (qPNL image) with reference to the converted quantization parameter (qPNL) derived for each block.

  Although the second type of input image data is described as two-dimensional data here, it may be three-dimensional data having width × height × number of channels (in this case, the number of channels may be 1). . The intermediate data on the neural network is also a “feature map” because it is data representing a certain kind of feature.

  Below, the specific example of the nonlinear transformation by the conversion part 402 is demonstrated.

(Nonlinear conversion example 1)
For example, the conversion unit 402 according to the present embodiment performs nonlinear conversion on an image having a pixel value as a quantization step applied when generating a decoded image (first type of input image data) supplied from the addition unit. It is generated as a subsequent qPNL image (converted second type input image data). The first type of input image data and the second type of input image data may be two-dimensional data having width × height, or may be three-dimensional data having width × height × number of channels. In the present embodiment, it is simply expressed as an image regardless of whether it is two-dimensional data or three-dimensional data. The quantization step may be a linearly converted value (a value obtained by multiplying and adding a certain value) as a value after conversion.

  As described above, the inverse quantization / inverse transform unit 311 according to the present embodiment obtains QP = 2 ^ (qP / 6) from the quantization parameter (qP) decoded with reference to the encoded stream Te. The quantization step to be calculated is calculated. Then, the quantization coefficient input from the entropy decoding unit 301 is inversely quantized using the quantization step, and further the inverse frequency transform is performed to calculate a residual signal. Further, a decoded image is generated by adding the calculated residual signal to the predicted image in the adding unit.

  In this example, the nonlinear transformation similar to the transformation from the quantization parameter to the quantization step used when generating the residual signal is used for the nonlinear transformation in the transformation unit 402.

  Thus, by using the quantization step used for quantization / inverse quantization when generating a decoded image for nonlinear transformation, the performance of the filter processing can be improved.

More specifically, the conversion unit 402 converts the quantization parameter (qP) into a nonlinear conversion quantization parameter (qPNL) using the following conversion formula. FIG. 6 is a graph showing the relationship between the quantization parameter (qP) and the transform quantization parameter (qPNL). As an example of conversion processing in decimal point arithmetic,
qPNL = 2 ^ (qP / 6) (Formula 1-1)
Can be used. FIG. 6A is a graph showing the relationship between the quantization parameter in the conversion equation of Expression 1-1 and the conversion quantization parameter.

As an example of integer arithmetic,
qPNL = levelScale [qP% 6] << (qP / 6) (Formula 1-2)
Can be used. The conversion expression of Expression 1-2 represents that a left bit shift of qP / 6 is performed on levelScale [qP% 6]. Here,% represents a remainder, and qP% 6 represents a remainder obtained by dividing qP by 6. levelScale [qP% 6] is an integer determined by the quantization parameter qP. As an example of the table levelScale [], for k = 0, 1, 2, 3, 4, 5,
levelScale [k] = {40, 45, 51, 57, 64, 72}
Can be used. FIG. 6B is a graph showing the relationship between the quantization parameter in Expression 1-2 and the transform quantization parameter.

  Note that the granularity of the quantization parameter qP is not limited to the above. That is, in the above description, a case is described in which the quantization step is doubled by a change of 6 in the quantization parameter qP. For example, the quantization step is doubled by a change in K of the quantization parameter qP. It may be the case. In this case, the outline of the operation is the same as described above. qPNL is derived as follows.

qPNL = 2 ^ (qP / K) (Formula 1-1 ')
qPNL = levelScale [qP% K] << (qP / K) (Formula 1-2 ')
In addition, FIG. 10 shows the result of nonlinear conversion of Formula 1-2. The quantization parameter (qP) shown in FIG. 10A is arranged for each unit region (or unit region having the same quantization parameter (qP)) of the transform unit. FIG. 10B shows the transform quantization parameter (qPNL) after the quantization parameter (qP) shown in FIG. 10A is nonlinearly transformed for each pixel.

(Nonlinear conversion example 2)
Also in this example, as in the above example, the conversion unit 402 performs nonlinear conversion on the quantization parameter (qP) given as the pixel value, and generates a conversion quantization parameter (qPNL) subjected to nonlinear conversion.

  However, in this example, the conversion unit 402 performs non-linear conversion on an image whose pixel value is the inverse of the quantization step applied to the decoded image (first type input image data) supplied from the addition unit. QPNL image (second type of input image data). Note that a linearly converted value (a value obtained by multiplying and adding the inverse of the quantization step) may be used as a value after conversion.

  Thus, the performance of the filtering process can also be improved by using the inverse of the quantization step when generating the decoded image.

More specifically, the conversion unit 402 converts the quantization parameter into a nonlinear conversion quantization parameter using the following conversion expression. FIG. 7 is a graph showing the relationship between the quantization parameter (qP) and the transform quantization parameter (qPNL). As an example of conversion processing in decimal point arithmetic,
qPNL = 1 / (2 ^ (qP / 6)) (Formula 2-1)
Can be used. FIG. 7A is a graph showing the relationship between the quantization parameter in the conversion equation of Expression 2-1 and the conversion quantization parameter.

As an example of integer arithmetic,
qPNL = quantScale [qP% 6] >> (qP / 6) (Formula 2-2)
Can be used. The conversion formula of Formula 2-2 represents that qP / 6 right bit shift is performed on quantScale [qP% 6]. Here, quantScale [qP% 6] is an integer determined by the quantization parameter qP. As an example of the table quantScale [], for k = 0, 1, 2, 3, 4, 5,
quantScale [k] = {26214, 23302, 20560, 18396, 16384, 14564}
Can be used. FIG. 7B is a graph showing the relationship between the quantization parameter in Expression 2-2 and the transform quantization parameter.

  The formula when the particle size of qP is K is as follows.

qPNL = 1 / (2 ^ (qP / K)) (Formula 2-1 ')
qPNL = quantScale [qP% K] >> (qP / K) (Formula 2-2 ')
In the following conversion examples, it is possible to change the granularity of qP to other than 6. However, in the case of the granularity K, only 6 is replaced with K, and the description is omitted.

(Nonlinear conversion example 3)
The transform unit 402 according to this example performs any one of logarithmic transformation, exponential transformation, and reciprocal transformation on each quantization parameter included as a pixel value of the qP image (second type of input image data). In other words, a qP image (second type of input image data) after nonlinear transformation is generated.

  More preferably, any one of logarithmic transformation, exponential transformation, and reciprocal transformation is used to transform the transformation quantization parameter so that the change width becomes smaller as the quantization parameter becomes larger.

  By performing the conversion as described above, the performance of the filter processing can be improved.

(Nonlinear conversion example 3-1)
As an example, the conversion unit 402 converts the quantization parameter into a nonlinear conversion quantization parameter using the following conversion formula. FIG. 8 shows a graph of the conversion formula. Examples of conversion formulas are:
qPNL = log (qP) (Formula 3-1)
Can be used. In other words, the transform quantization parameter is a logarithmic transform of the quantization parameter. Here, as the base of the logarithm, for example, 2 can be used, but this does not limit the present example.

  FIG. 8A is a graph showing the relationship between the quantization parameter in the conversion equation of Expression 3-1 and the conversion quantization parameter.

In addition, when qP = 0, in order to avoid the divergence of log (qP), c may be added to the logarithm as an offset and processed as follows.
qPNL = k * log (qP + c) (Formula 3-2)
Here, k represents a predetermined normalization coefficient. However, the specific value of k does not limit the present embodiment.

(Nonlinear Conversion Example 3-2)
Moreover, the conversion part 402 may convert a quantization parameter into a nonlinear conversion quantization parameter using the following conversion formulas. Examples of conversion formulas are:
qPNL = (qP / 6) ^ r (where 0 <r <1, for example, r = 1/2) (Equation 3-3)
Is mentioned. In other words, the transform quantization parameter is an exponential transform of the quantization parameter. FIG. 8B is a graph showing the relationship between the quantization parameter in Expression 3-3 and the transform quantization parameter.

As in the logarithmic conversion example, the offset c may be added and the following processing may be performed.
qPNL = k * ((qP + c) / 6)) ^ 1/2 (Formula 3-4)
(Nonlinear conversion example 3-3)
Moreover, the conversion part 402 may convert a quantization parameter into a nonlinear conversion quantization parameter using the following conversion formulas. Examples of conversion formulas are:
qPNL = 1 / qP (Equation 3-5)
Is mentioned. In other words, the transform quantization parameter is an inverse transform of the quantization parameter. FIG. 8C is a graph showing the relationship between the quantization parameter in Expression 3-5 and the transform quantization parameter.
As in the above example, the offset c may be added and processed as follows.

qPNL = k / (qP + c) (Formula 3-6)
Note that a predetermined value can be used as the value of the offset c in the above example, but this does not limit the present embodiment. For example, as described below, the offset c may be decoded from the encoded stream Te and used.

(Offset of quantization coefficient qP)
The value of the offset c in the above example may be determined as follows. That is, a syntax for specifying the value of the offset c may be included in the slice header, and the entropy decoding unit 301 may decode the syntax to derive the offset c.

  More specifically, the syntax slice_beta_offset_div2 for specifying the offset is included in the slice header of the encoded stream Te, and the entropy decoding unit 301 decodes the syntax and then uses the following conversion formula to perform the offset. c may be derived.

c = slice_beta_offset_div2 << 1 (Formula 3-7)
Note that slice_beta_offset_div2 is also a syntax element for adjusting the ease of deblocking filter processing.

(Clip processing)
Moreover, you may use each example shown to the above-mentioned nonlinear conversion examples 1-3 in combination with a clip process.

For example, after adding the offset derived using Equation 3-7 to qP, clipping to a predetermined range, and substituting qP after the clipping into Equations 3-1, 3-3, 3-5, etc. May be. As an example, when clipping to the range of 0 to 51,
qP = Clip3 (0, 51, qP + (slice_beta_offset_div2 << 1)) (Equation 3-8)
QP derived by the above may be substituted into Equations 3-1, 3-3, 3-5, and the like.

(Use of boundary strength)
Further, the value of the offset c may be determined according to the boundary strength (bS). For example, the offset c may be derived using the following conversion formula.

c = 2 * (bS-1) + (slice_tc_offset_div2 << 1) (Equation 3-9)
Here, bS is a parameter for adjusting the strength of the deblocking filter processing determined for each pixel, and for example, according to the block including the pixel and each adjacent block adjacent to the block, the following (a) It is determined as (c).
(a) bS = 2
When at least one block is an intra prediction block.
(b) bS = 1
-When a non-zero orthogonal transform coefficient exists in at least one block and is a TU boundary.

  When the absolute value of the motion vector difference between two blocks that sandwich the boundary is 1 pixel or more.

When the motion compensation reference images of the two blocks sandwiching the boundary are different or the number of motion vectors is different.
(c) bS = 0
When other than (a) and (b) above.

  Also, the syntax slice_tc_offset_div2 in Expression 3-9 may be included in the slice header of the encoded stream Te as an example of the syntax for specifying the offset, and may be decoded by the entropy decoding unit 301.

  Note that slice_tc_offset_div2 is also a syntax element for determining whether the deblocking filter processing mode is a strong filter or a weak filter and adjusting the filter strength.

Further, after adding the offset derived using Equation 3-9 to the quantization parameter (qP), the clip is clipped to a predetermined range in the same manner as the above (Clip processing), and the quantization parameter after the clip (QP) may be substituted into Equations 3-1, 3-3, 3-5, and the like. As an example, when clipping to the range of 0 to 51,
qP = Clip3 (0, 51, qP + 2 * (bS-1) + (slice_tc_offset_div2 << 1)) (Equation 3-10)
QP derived by the above may be substituted into Equations 3-1, 3-3, 3-5, and the like.

(Nonlinear conversion example 4)
In addition, the transform unit 402 according to the present embodiment performs nonlinear transform using a polynomial on the quantization parameter applied to the image after nonlinear transformation, for example, thereby performing an image after nonlinear transformation (the second type of image). Input image data).

  In this example, as an example, the nonlinear transformation can be performed using a polynomial that approximates at least one of the logarithmic transformation, the exponential transformation, and the inverse transformation described in the nonlinear transformation example 3.

  As described above, by applying nonlinear transformation using polynomial approximation to the quantization parameter, the parameter can be converted into a parameter suitable for the filter processing, and the performance of the filter processing can be improved.

As an example, the conversion unit 402 converts the quantization parameter (qP) into a nonlinear conversion quantization parameter (qPNL) using the following conversion formula. Examples of conversion formulas are:
qPNL = a * qP + b * qP ^ 2 + c (Formula 4-1)
Can be used.

The order of the polynomial is not limited to the second order, and a polynomial approximation of the third order or higher may be performed. For example, as a cubic polynomial:
qPNL = a * qP + b * qP ^ 2 + c * qP ^ 3 + d (Formula 4-2)
May be used.

  Further, the coefficients (a, b, c) of the equation 4-1 and the coefficients (a, b, c, d) of the equation 4-2 may use predetermined values, and the syntax corresponding to the coefficients. May be included in the encoded stream Te.

  Further, the type and number of images input to the neural network unit 404 are not limited. For example, one or more additional images may be input to the neural network unit 404 in addition to the decoded image generated by the adding unit and the non-linearly converted image. As an example, as shown in FIG. 9A, in addition to the decoded image (first type input image data) generated by the adding unit and the non-linearly converted qPNL image (converted second type input image data) The qP image before nonlinear conversion may be input to the neural network unit 404. As shown in FIG. 9B, a plurality of qPNL images after nonlinear transformation may be input to the neural network unit 404 together with the qP image before nonlinear transformation. The plurality of qPNL images after nonlinear transformation may be transformed by different nonlinear transformation methods, but it is preferable to input images complementary to each other. For example, a qPNL image after nonlinear conversion using Equation 1-1 is input as the qPNL image after the first nonlinear transformation, and nonlinear transformation using Equation 2-1 as the qPNL image after the second nonlinear transformation. A configuration for inputting a later qPNL image is given.

(Quantization parameter (qP))
The quantization parameter (qP) described above is a parameter that controls the compression rate and image quality of an image. In this embodiment, the quantization parameter (qP) has a characteristic that the larger the value, the lower the image quality and the smaller the code amount, and the smaller the value, the higher the image quality and the larger the code amount. As the quantization parameter (qP), for example, a parameter for deriving the quantization width of the transform transform coefficient related to the prediction residual can be used.

  As the quantization parameter (qP) for each picture, one representative quantization parameter (qP) of the processing target frame can be input. For example, the quantization parameter (qP) can be specified by a parameter set applied to the current picture. Also, the quantization parameter (qP) can be calculated based on the quantization parameter (qP) applied to the picture component. Specifically, the quantization parameter (qP) can be calculated based on the leading value or average value of the quantization parameter (qP) applied to the slice.

  Further, as a quantization parameter (qP) for a unit obtained by dividing a picture, a quantization parameter (qP) for each unit obtained by dividing a picture on a predetermined basis can be input. For example, the quantization parameter (qP) can be applied for each slice. Also, the quantization parameter (qP) can be applied to the blocks in the slice. In addition, the quantization parameter (qP) can be specified in a region unit independent of existing coding units (for example, each region obtained by dividing a picture into 16 × 9). In this case, the quantization parameter (qP) depends on the number of slices and the number of transform units. As a result, the value of the quantization parameter (qP) corresponding to the region becomes indefinite, and the CNN filter cannot be configured. Therefore, a method using the average value of the quantization parameter (qP) in the region as a representative value can be considered. There is also a method using the quantization parameter (qP) at one position in the region as a representative value. There is also a method of using the median value or mode value (mode) of quantization parameters (qP) at a plurality of positions in the region as representative values.

  Also, when inputting a specific number of quantization parameters (qP), a list of quantization parameters (qP) is generated and input to the CNN filter so that the number of quantization parameters (qP) is constant. May be. For example, a method of creating a list of quantization parameters (qP) for each slice, and creating and inputting a list of three quantization parameters (qP) of a maximum value, a minimum value, and a median value is conceivable.

  In addition, as the quantization parameter (qP) for each component, a quantization parameter (qP) applied to the component to be processed can be input. Examples of the quantization parameter (qP) include a luminance quantization parameter (qP) and a color difference quantization parameter (qP).

  In addition, when the CNN filter is applied in block units, the quantization parameter (qP) of the target block and the quantization parameter (qPNeighbour) around the block may be input as the peripheral quantization parameter (qP).

  The CNN filter 305 can be designed according to a picture and a coding parameter other than the quantization parameter (qP). That is, since the CNN filter 305 can be designed according to not only picture characteristics that can be derived from image data such as directionality and activity, but also encoding parameters, the CNN filter 305 has different strengths for each encoding parameter. This filter can be realized. Therefore, since the present embodiment includes the CNN filter 305, it is possible to perform processing according to the encoding parameter without introducing a different network for each encoding parameter.

<Second Embodiment>
As in the first embodiment, the image filter device according to the present embodiment includes a qP image generation unit 400, a neural network unit 404, and a conversion unit 402, as shown in FIG.

  Since the function of the qP image generation unit 400 is the same as that of the qP image generation unit 400 described in the first embodiment, detailed description thereof is omitted.

  The conversion unit 402 according to the present embodiment performs, for example, conversion that standardizes (normalizes) a quantization parameter that is a pixel value of a qP image (second type of input image data), thereby converting a converted qPNL image ( The second type of input image data after conversion) is generated.

  The post-conversion 402 performs conversion after normalizing each quantization parameter (qP) included in the qP image acquired from the qP image generation unit 400 to −1 to 1 or a range included in the quantization parameter. q Generate a PNL image. The standardized quantization parameter value is called a standard quantization parameter (qPNL0).

  For example, the transform unit 402 derives a quantization parameter qPNL after nonlinear transformation from the quantization parameter (qP) acquired from the qP image generation unit 400, and further performs standard quantization from the quantization parameter qPNL after nonlinear transformation. By deriving the parameter (qPNL0), a qPNL image after conversion / standardization using the standard quantization parameter (qPNL0) as each pixel value is generated. For the nonlinear conversion, the nonlinear conversion example described in <First Embodiment> may be used.

  Hereinafter, a specific example will be described in which the transform unit 402 performs nonlinear transform on the quantization parameter (qP) and derives the standard quantization parameter (qPNL0) from the quantized parameter (qPNL) after the nonlinear transform.

(Subtract fixed value)
A standard quantization parameter (qPNL0) is derived by subtracting a specific QP value (modQP) from the quantization parameter (qP).

As an example of a specific formula,
qPNL0 = (qP-modQP) / v (Formula 5-1)
Can be used.

  modQP is a value of a quantization parameter having a high occurrence frequency. Here, when the value of the quantization parameter having a high occurrence frequency among the values of the quantization parameter is 32, the standard quantization parameter (qPNL0) can be obtained with modQP = 32.

Moreover, v in Formula 5-1 is calculated | required from the following formula.
v = max ((maxQP-modQP), (modQP-minQP)) (Formula 5-2)
maxQP is the maximum value (for example, 51) that the quantization parameter can take, and minQP is the minimum value (for example, 0) that the quantization parameter can take.

  In addition, the transforming unit 402 performs nonlinear transformation on the quantization parameter (qP) and the quantization parameter (modQP) having a high occurrence frequency, and then determines the occurrence frequency after the nonlinear transformation from the quantization parameter NLfunc (qP) after the nonlinear transformation. The standard quantization parameter (qPNL0) may be derived by subtracting the high quantization parameter NLfunc (modQP).

As an example of a specific formula,
qPNL0 = (NLfunc (qP)-NLfunc (modQP)) / v (Formula 5-3)
Can be used. In the above description, NLfunc is a function for performing non-linear conversion, and specifically, various non-linear conversion formulas described in the first embodiment can be used.

Here, v in Expression 5-3 is obtained from the following expression as an example.
v = max {(NLfunc (maxQP) -NLfunc (modQP)), (NLfunc (modQP) -NLfunc (minQP))} (Formula 5-4)
(Clip processing)
In addition, the conversion unit 402 according to the present embodiment may derive the standard quantization parameter (qPNL0) by performing clipping processing on the quantization parameter (qP) within a predetermined range. The derived standard quantization parameter (qPNL0) may be input to the neural network unit 404.

  Here, the predetermined range may be configured to specify a parameter value range, and the specified range may be used as the predetermined range, or a predetermined range may be used as the predetermined range.

As an example of a specific formula,
qPNL0 = Clip3 (minDNQP, maxDNQP, qP) (Formula 6-1)
Can be used.
minDNQP represents the minimum value of the quantization parameter (qP) in the predetermined range, and maxDNQP represents the maximum value of the quantization parameter (qP) in the predetermined range, and minDNQP <maxDNQP. As an example, minDNQP = 12, maxDNQP = 51 can be used.

  Further, the transform unit 402 performs non-linear transformation on the quantization parameter (qP), and then performs clipping processing on the transformed quantization parameter (qPNL) obtained by nonlinear transformation within a predetermined range, thereby deriving the standard quantization parameter (qPNL0). Then, it may be input to the neural network unit 404.

As an example of a specific formula,
qPNL0 = Clip3 (minDNQPNL, maxDNQPNL, qPNL) (Formula 6-2)
Can be used. minDNQPNL represents the minimum value of the transform quantization parameter (qPNL) in the predetermined range, and maxDNQPNL represents the maximum value of the transform quantization parameter (qPNL) in the predetermined range, and minDNQPNL <maxDNQPNL.

<Third Embodiment>
As in the first embodiment, the image filter device according to the present embodiment includes a qP image generation unit 400, a neural network unit 404, and a conversion unit 402, as shown in FIG.

  Since the function of the qP image generation unit 400 is the same as that of the qP image generation unit 400 described in the first embodiment, detailed description thereof is omitted.

  For example, the conversion unit 402 according to the present embodiment refers to an encoding parameter for generating a prediction image or a difference image (not shown) and corrects a qP image (second type input image data). By converting, a converted qPNL image (converted second type input image data) is generated.

  The conversion unit 402 generates a converted qPNL image by converting each quantization parameter (qP) included in the qP image acquired from the qP image generation unit 400. Here, a value obtained by correcting the quantization parameter (qP) with reference to the encoding parameter is referred to as a modified quantization parameter (qP1). The transform unit 402 according to the third embodiment derives a modified quantization parameter (qP1) from the quantization parameter (qP) acquired from the qP image generation unit 400 with reference to the coding parameter, and the modified quantization parameter ( A qPNL image after conversion is generated by performing nonlinear conversion on qP1). Nonlinear transformation is the same as described above.

  Hereinafter, a specific example in which the transform unit 402 derives the modified quantization parameter (qP1) from the quantization parameter (qP) with reference to the encoding parameter will be described.

((Modification for all pixels in the block))
In the following A1 to A6, an example in which the quantization parameters of all pixels included in the target block are corrected will be described.

  However, the following example may not be applied to all the pixels of the target block, but may be applied only to the boundary neighboring pixels described later.

(A1: Correction based on inter and intra)
As an example, the conversion unit 402 refers to the encoding parameter supplied from the prediction parameter decoding unit, and determines whether the target block is a block to which intra prediction is applied or not. When it is determined that the target block is a block to which intra prediction is applied, the transform unit 402 increases the quantization parameter (qP) for all the pixels of the block, and derives a modified quantization parameter (qP1). . A prediction block to which intra prediction is applied is also referred to as an intra prediction block. A block other than the intra prediction block is a prediction block to which inter prediction is applied.

  Thus, the performance of the filter processing can be improved by increasing the quantization parameter (qP) in the intra prediction block.

As an example, the transform unit 402 derives a modified quantization parameter (qP1) from the quantization parameter (qP) using the following equation. Examples of expressions include
qP1 = qP + (intraPred? offset: 0) (Formula 7-1)
Can be used. IntraPred in Equation 7-1 is a Boolean expression that asks whether the target block is an intra prediction block. If true, that is, if the target block is an intra prediction block, the quantization parameter (qP) Offset is added to. Further, if it is false, that is, if the target block is other than the intra prediction block, nothing is added. As the offset, a predetermined value may be used, or the encoded stream Te may be included in the SPS or PPS and notified.

(A2: Correction based on luminance and color difference)
In addition, the conversion unit 402 according to the present embodiment determines whether the target block is a luminance block or a color difference block. After the determination, when the target block is a color difference block, the conversion unit 402 increases the quantization parameter (qP) of the block, and derives a modified quantization parameter (qP1).

  Thus, by increasing the quantization parameter (qP) in the color difference block, the performance of the filter processing can be improved.

As an example, the transform unit 402 derives a modified quantization parameter (qP1) from the quantization parameter (qP) using the following equation. Examples of expressions are:
qP1 = qP + (cIdx! = 0)? offset: 0 ... (Formula 8-1)
・ In the case of luminance block cIdx = 0 (Formula 8-2)
-In case of color difference block cIdx = 1, 2 (Formula 8-3)
Can be used. In Expression 8-1, when cIdx is not 0, it is determined to be true, offset (offset) is added to qP, and when cIdx is 0, it is determined to be false, and nothing is added to qP. As the offset, a predetermined value may be used, or the encoded stream Te may be included in the SPS or PPS and notified. Note that cIdx in Equation 8-1 is 0 for the luminance block as shown in Equation 8-2 and 1 or 2 for the color difference block as shown in Equation 8-3.

(A3: Correction based on motion vector determination)
(Motion vector determination 1-1)
When the transform unit 402 according to the present embodiment determines that the difference between the motion vectors of the target block and the adjacent block is larger than a certain value, the transform unit 402 modifies the modified quantization parameter from the quantization parameter (qP) regarding all the pixels included in the target block. (QP1) may be derived.

  That is, when the difference between the motion vector of block P (mvp [0], mvp [1]) and the motion vector of block Q adjacent to block P (mvq [0], mvq [1]) is greater than a predetermined value In addition, mod1 may be added to the quantization parameter (qP) for all the pixels of the target block.

As an example of a specific judgment formula,
abs (mvp [0]-mvq [0]) + abs (mvp [1]-mvq [1])> diff (Equation 8a-1)
Is mentioned. A predetermined value may be used for diff. When abs (mvp [0] −mvq [0]) + abs (mvp [1] −mvq [1]) is larger than diff in the determination formula of Formula 8a-1, for example, the following formula:
qP1 = qP + mod1 (Formula 8a-2)
Thus, mod1 may be added to derive the modified quantization parameter (qP1) from the quantization parameter (qP). mod1 is a predetermined value.

(Motion vector determination 1-2)
In addition, the conversion unit 402 according to the present embodiment, when any of the differences between the horizontal component and the vertical component of the motion vector of the target block and the adjacent block is larger than a predetermined value, all the pixels included in the target block A predetermined mod1 may be added to the quantization parameter (qP) for.

As an example of a specific judgment formula,
(abs (mvp [0]-mvq [0])> diff) || (abs (mvp [1]-mvq [1])> diff) (Formula 8b-1)
Is mentioned. Here, abs (a) is a function for deriving the absolute value of a.

  In Formula 8b-1, when any one of abs (mvp [0] -mvq [0]) and abs (mvp [1] -mvq [1]) is greater than diff, for example, Formula 8a-2 The modified quantization parameter (qP1) may be derived from the quantization parameter (qP) using the conversion equation.

(A4: Modification based on reference picture index)
The transform unit 402 refers to the reference picture index included in the encoding parameter, and the values of the reference picture index associated with the target block P and the reference picture index associated with the adjacent block Q adjacent to the target block P Are different from each other with respect to the quantization parameter (qP) of each pixel included in the target block, for example, using the transformation formula of Expression 8a-2, the modified quantization parameter (qP1) May be derived.

(A5: Correction based on conversion coefficient)
Also, the transform unit 402 according to the present embodiment refers to the coding parameter supplied from the prediction parameter decoding unit, and adds an offset to the quantization parameter (qP) when the target block has a non-zero transform coefficient. May be. The quantization parameter (qP) to which the offset is added is used as the modified quantization parameter (qP1).

  As described above, when the target block has a transform coefficient, the performance of the filter processing can be improved by increasing the quantization parameter (qP).

As an example, the transform unit 402 derives a modified quantization parameter (qP1) from the quantization parameter (qP) using the following equation. Examples of expressions are:
qP1 = qP + (cbp! = 0)? offset: 0 (Equation 9-1)
-When the conversion coefficient is included in the target block cbp = 1 (Equation 9-2)
-When the conversion coefficient is not included in the target block cbp = 0 (Equation 9-3)
Can be used. Here, cbp indicates a coding block pattern (cbf_luma, cbf_cb, cbf_cr). Expression 9-1 is determined to be true when cbp is not 0, offset (offset) is added to qP, is determined to be false when cbp is 0, and nothing is added to qP. As the offset, a predetermined value may be used, or the encoded stream Te may be included in the SPS or PPS and notified. Note that cbp in Equation 9-1 is 1 when the conversion coefficient is included in the target block as shown in Equation 9-2, and no conversion coefficient is included in the target block as shown in Equation 9-3. The case is 0.

(Correction based on block size)
In addition, the conversion unit 402 according to the present embodiment may derive the modified quantization parameter (qP1) by adjusting the size of the quantization parameter (qP) according to the block size of the target block.

  For example, when the block size of the target block is larger than a predetermined threshold size, a modified quantization parameter (qP1) may be derived by adding a predetermined value to the quantization parameter (qP).

(Derivation example 1)
As an example, the transform unit 402 derives a modified quantization parameter (qP1) from the quantization parameter (qP) using the following equation. Examples of expressions are:
qP1 = (blkW + blkH> TH)? qP + mod: qP (Equation 10-1)
Can be used. Here, blkW in Equation 10-1 is the width of the target block, blkH is the height of the target block, and TH is the size of a predetermined threshold. When the sum of the width and height of the target block is larger than a predetermined threshold, it is determined to be true, a predetermined mod is added to the quantization parameter (qP), and a modified quantization parameter (qP1) is derived. . When the sum of the width and height of the target block is smaller than a predetermined value, it is determined to be false, and nothing is added to the quantization parameter (qP).

Further, the modified quantization parameter (qP1) may be derived by adjusting the value added to the quantization parameter (qP) according to the block size. For example, two predetermined threshold values for the block size may be set, and a value to be added to the quantization parameter (qP) may be set according to each threshold value. As an example of the formula when two predetermined thresholds are provided,
if (blkW + blkH> TH1)
qP1 = qP + mod1;
else if (blkW + blkH> TH2)
qP = qP + mod2; (Formula 10-2)
Can be used. Here, TH1 and TH2 are predetermined threshold values (TH1> TH2), respectively. First, when the sum of the width and height of the target block is larger than a predetermined threshold TH1, a predetermined mod1 is added to the quantization parameter (qP) to derive a modified quantization parameter (qP1). Otherwise, the expression after else if is executed. When the sum of the width and height of the target block is larger than the threshold value TH2, a predetermined mod2 is added to the quantization parameter (qP) to derive a modified quantization parameter (qP1). In Expression 10-2, two threshold values are set, but the number of threshold values is not particularly limited, and three or more threshold values may be set. For mod, mod1, and mod2, predetermined values may be used, or the encoded stream Te may be included in the SPS or PPS for notification.

(Derivation example 2)
As an example, the conversion unit 402 may derive and add an adjustment term from a value referred to by a lookup table according to the block size of the target block.

Examples of expressions include
qP1 = qP + qPmodTable [log2 (blkW) + log2 (blkH)] (Formula 11-1)
Can be used. Here, blkW is the width of the target block, and blkH is the height of the target block.

(Derivation Example 3)
As an example, the conversion unit 402 may derive and add an adjustment term to the block size of the target block by product, sum, or shift operation.

Examples of expressions include
qP1 = qP + ((a * (log2 (blkW) + log2 (blkH))) >> b) + c (Equation 12-1)
Can be used. Here, a, b, and c are predetermined values.

(A6: Correction based on block depth)
In addition, the transform unit 402 according to the present embodiment refers to the coding parameter depth information (division information), changes the size of the quantization parameter (qP) according to the depth of the target block, and corrects it. The quantization parameter (qP1) may be derived.

  The block depth is a value incremented every time a block is divided by block division (quadrant tree QT division, binary tree BT division). The depth of the block increases each time the block is divided with the highest level being 0.

((Correction to pixels near block boundary))
In the following B1 to B4, an example in which the quantization parameter of the neighboring pixel at the boundary between the target block and the adjacent block adjacent to the target block will be described.

  The following example may be applied not only to the pixels near the boundary but also to all the pixels of the target block.

(B1: Block boundary correction)
The transform unit 402 according to the present embodiment determines whether or not the block boundary requires a modification of the quantization parameter (qP), and in the case of the block boundary, the distance from the block boundary is within a predetermined number of pixels. The quantization parameter (qP) of the pixel (near the boundary, similar in this specification) may be increased to derive the modified quantization parameter (qP1). Here, the predetermined number of pixels may be one or two or more. The case where the predetermined number of pixels is 1 corresponds to a pixel adjacent to the boundary.

  As described above, the performance of the filter processing can be improved by increasing the quantization parameter (qP) of the pixels near the boundary.

For example, the quantization parameter (qP) of the pixel near the boundary can be corrected using the following equation. As an example of the expression when the block boundary of the target block is a horizontal boundary,
if (horizontal boundary) {// upper boundary or lower boundary
for (x = 0; x <blkW; x ++) {
qP1 [x] [0] = qP [x] [0] + mod
qP1 [x] [blkH-1] = qP [x] [blkH-1] + mod
}
} (Formula 13-1)
Can be used. blkW is the number of pixels in the horizontal direction of the block, blkH is the number of pixels in the vertical direction of the block, and [x] [y] are coordinates (x, y) corresponding to the pixels of the block.

In addition, as an example of an expression when the block boundary of the target block is not a horizontal boundary but a vertical boundary,
if (vertical border) {// left border or right border
for (y = 0; y <blkH; y ++) {
qP1 [0] [y] = qP [0] [y] + mod
qP1 [blkW-1] [y] = qP [blkW-1] [y] + mod
}
} (Formula 13-2)
Can be used.

  FIG. 19 shows a specific example of pixels near the boundary. FIG. 19A shows the quantization parameter of the target block before correction when the predetermined number of pixels is 1, and FIG. 19B shows the quantization parameter of the target block after correction. For example, when mod = 2, 2 is added to the quantization parameter of 1 row and 1 column adjacent to the boundary as shown in FIG.

  Similarly, FIG. 20A shows the quantization parameter of the target block before correction when the predetermined number of pixels is 2, and FIG. 20B shows the quantization parameter of the target block after correction. For example, when mod = 2, 2 is added to the 2 × 2 quantization parameter adjacent to the boundary as shown in FIG.

  In addition, when correcting the N pixel at the block boundary, the conversion unit 402 according to the present embodiment may increase the correction amount of the quantization parameter (qP) as the pixel is closer to the boundary line.

As an example of a specific formula,
if (horizontal boundary) {// upper boundary or lower boundary
for (x = 0; x <blkW; x ++) {
qP1 [x] [0] = qP [x] [0] + mod1
qP1 [x] [1] = qP [x] [1] + mod2
qP1 [x] [blkH-1] = qP [x] [blkH-1] + mod1
qP1 [x] [blkH-2] = qP [x] [blkH-2] + mod2
}
} (Formula 14-1)
else if (vertical border) {// left border or right border
for (y = 0; y <blkH; y ++) {
qP1 [0] [y] = qP [0] [y] + mod1
qP1 [1] [y] = qP [1] [y] + mod2
qP1 [blkW-1] [y] = qP [blkW-1] [y] + mod1
qP1 [blkW-2] [y] = qP [blkW-2] [y] + mod2
}
} (Formula 14-2)
Can be used. In Expressions 14-1 and 14-2, for example, mod1 = 2 and mod2 = 1 may be set. FIG. 21A shows the quantization parameter of the target block before correction when the predetermined number of pixels is 2, and FIG. 21B uses the expressions 14-1 and 14-2 to calculate the pixels near the boundary. The quantization parameter of the target block after deriving the modified quantization parameter (qP1) from the quantization parameter (qP) is shown.

(Block boundary determination)
Specifically, the determination as to whether or not the block boundary requires the modification of the quantization parameter (qP) may be performed using any of the following determination methods. In two adjacent blocks,
・ If the quantization parameter (qP) is different,
・ If intra prediction mode is used for at least one
If there is a non-zero transform coefficient in at least one
-In the inter prediction mode, if the motion vector is different,
-In inter prediction mode, if the reference picture is different,
It is. For example, FIG. 23A shows the quantization parameters p0 and q0 of adjacent blocks P and Q at the horizontal boundary. FIG. 23B shows the quantization parameters p0 and q0 of adjacent blocks P and Q at the vertical boundary.

As shown in FIG. 22 (a), even when there are a plurality of boundary lines in one target block and each block is further divided into a plurality of blocks, each block is determined using the above-described (block boundary determination). The boundary can be determined. After the determination of the block boundary, in each divided block, the modified quantization parameter (qP1) is calculated from the quantization parameter (qP) of the pixel adjacent to each boundary line using Expression 13-1 and Expression 13-2. It may be derived. FIG. 22B shows the target block after correction. In this case, mod = 2 in Expression 13-1 and Expression 13-2, and one row and one column (predetermined number of pixels is adjacent to each boundary line). 2 is added to the quantization parameter of 1).

(B2: Correction based on motion vector determination)
(Motion vector determination 1-1)
When the transform unit 402 according to this embodiment determines that the difference between motion vectors of adjacent blocks is larger than a certain value, the transform unit 402 calculates the modified quantization parameter (qP1) from the quantization parameter (qP) of the target block related to the boundary neighboring pixels. It may be derived.

  That is, when the difference between the motion vector of block P (mvp [0], mvp [1]) and the motion vector of block Q adjacent to block P (mvq [0], mvq [1]) is greater than a predetermined value In addition, mod1 may be added to the quantization parameter (qP) for the pixels near the boundary.

As an example of a specific judgment formula,
abs (mvp [0]-mvq [0]) + abs (mvp [1]-mvq [1])> diff (Equation 15-1)
Is mentioned. A predetermined value may be used for diff. When abs (mvp [0] −mvq [0]) + abs (mvp [1] −mvq [1]) is larger than diff in the determination formula of Formula 15-1, for example, the following formula:
qP1 = qP + mod1 (Formula 15-2)
Thus, mod1 may be added to derive the modified quantization parameter (qP1) from the quantization parameter (qP). mod1 is a predetermined value.

(Motion vector determination 1-2)
In addition, the transform unit 402 according to the present embodiment performs a quantization parameter on the boundary neighboring pixels when one of the differences between the horizontal components and the vertical components of the motion vectors of adjacent blocks is larger than a predetermined value. A predetermined mod1 may be added to (qP).

As an example of a specific judgment formula,
(abs (mvp [0]-mvq [0])> diff) || (abs (mvp [1]-mvq [1])> diff) (Equation 16-1)
Is mentioned. Here, abs (a) is a function for deriving the absolute value of a.

  In Expression 16-1, when any one of abs (mvp [0] -mvq [0]) and abs (mvp [1] -mvq [1]) is larger than diff, for example, Expression 15-2 The modified quantization parameter (qP1) may be derived from the quantization parameter (qP) related to the pixels near the boundary using the conversion formula.

(B3: Correction based on prediction mode determination)
Also, the conversion unit 402 according to the present embodiment refers to the encoding parameter, determines whether the prediction mode predModeP of the block P and the prediction mode predModeQ of the block Q adjacent to the block P are equal, and is not equal In this case, a predetermined mod1 may be added to the quantization parameter (qP) related to the boundary neighboring pixels.

As an example of a specific judgment formula,
predModeP! = predModeQ (Equation 17-1)
Is mentioned.

  In Formula 17-1, when the prediction mode predModeP and the prediction mode predModeQ are not equal, for example, the modified quantization parameter (qP1) is derived from the quantization parameter (qP) using the formula of Formula 15-2. May be.

(B4: Correction based on division depth determination)
In addition, the conversion unit 402 according to the present embodiment refers to the encoding parameter, determines whether or not the division depth depthP of the block P is equal to the division depth depthQ of the block Q adjacent to the block P, and is not equal. In this case, a predetermined mod1 may be added to the quantization parameter (qP) for the pixels near the boundary.

As an example of a specific judgment formula,
depthP! = depthQ (Formula 18-1)
Is mentioned.

  In Expression 18-1, when the divided depth depthP and the divided depth depthQ are not equal, for example, the modified quantization parameter (qP1) is derived from the quantization parameter (qP) using the expression of Expression 15-2. May be.

<Fourth Embodiment>
Another embodiment of the present invention is described below with reference to FIG. For convenience of explanation, description of members having the same functions as those described in the embodiment is omitted.

  In the present embodiment, as shown in FIG. 11, the CNN filter 305d includes a first-stage CNN filter 305d1 that is a plurality of individual neural networks including n + 1 CNN filters CNN0, CNN1,. 305d2, a second-stage CNN filter 305d3 that is a common neural network, a qP image generation unit 400d, and a conversion unit 402d.

  The conversion unit 402d converts the quantization parameter using any one of the above-described first to third embodiments, and supplies a representative value of the converted quantization parameter (qPNL) to the selector 305d2.

  In the first-stage CNN filter 305d1, for example, the CNN filter CNN0 includes three channels including a luminance (Y) channel, a first color difference (Cb) channel, and a second color difference (Cr) channel. The pre-filter image composed of is input. The CNN filter CNN0 is a filter optimized for the filter parameter FP having a value smaller than FP1. The CNN filter CNN1 is a filter optimized for a filter parameter FP having a value that is greater than or equal to FP1 and smaller than FP2. The CNN filter CNNn is a filter optimized for the filter parameter FP having a value equal to or greater than FPn.

  Each of the CNN filters CNN0, CNN1,..., CNNn included in the first-stage CNN filter 305d1 outputs the filtered image to the selector 305d2. The filter parameter FP is input to the selector 305d2, and an image that has been subjected to filter processing to be output to the second-stage CNN filter 305d3 is selected according to the input filter parameter FP. Therefore, the image that has been subjected to the optimum filter processing selected by the selector 305d2 is input to the second-stage CNN filter 305d3. In other words, the individual neural network in this embodiment selectively acts on the input image data according to the value of the filter parameter in the input image data to the image filter device.

  Note that the filter parameter FP for selecting the CNN filter may be explicitly encoded in the encoded data, or may be derived from the encoding parameter. For example, the filter parameter FP may be derived from a representative value (such as an average value) of a quantization parameter that is one of the encoding parameters.

  The second-stage CNN filter 305d3 performs a filtering process on the input image and outputs a filtered image. In other words, the common neural network in this embodiment acts in common on the output image data of the individual neural network regardless of the value of the filter parameter.

  The filter parameter FP used for selection by the selector 305d2 is not limited to the representative value of the converted quantization parameter (qPNL). The filter parameter FP may be explicitly transmitted in the encoded data. Further, as the filter parameter FP, in addition to the quantization parameter in the input image, a parameter indicating the type of intra prediction and inter prediction in the input image, a parameter indicating the intra prediction direction in the input image, and the partition depth of the input image (depth) Information, division information), and a parameter indicating the size of the partition in the input image. In these parameters, representative values such as a specific position value (upper left or center), average value, minimum value, maximum value, median value, and mode value may be used.

  FIG. 12 is a schematic diagram illustrating a modification of the configuration of the image filter device according to the present embodiment. As shown in FIG. 12, the CNN filter 305e includes n + 1 CNN filters CNN0, CNN1,..., CNNn, and the CNN filter 305e2 that is a plurality of individual neural networks, and the selector 305e3 are CNN filters that are common neural networks. It can also be arranged after 305e1. In this case, the CNN filter 305e1 acts on the input image data to the CNN filter 305e, and the CNN filter 305e2 selectively selects the output image data of the CNN filter 305e1 according to the value of the filter parameter in the input image data. Works. The selector 305e3 outputs the filtered image. Note that the CNN filter 305e includes a qP image generation unit 400e and a conversion unit 402e, which have the same configurations as the qP image generation unit 400d and the conversion unit 402d, respectively.

  Also in this embodiment, the CNN filter in the image decoding apparatus has the same function as the CNN filter in the image encoding apparatus.

  According to the configuration of the fourth embodiment, by using a part (305e2) that switches the network according to the size of the filter parameter FP and a part (305e1) that uses the same network regardless of the size of the filter parameter. Thus, the network configuration can be made smaller than the configuration in which the entire filter is switched with an encoding parameter such as a quantization parameter. The smaller the network configuration, the smaller the calculation amount and the higher the speed. In addition, there is an effect that the learning parameter is robust and appropriate filter processing can be performed for many input images.

<Fifth Embodiment>
The following will describe another embodiment of the present invention with reference to FIG. For convenience of explanation, description of members having the same functions as those described in the embodiment is omitted.

  In the present embodiment, as shown in FIG. 13, the CNN filter 305f includes a first stage CNN filter 305f1 including n + 1 CNN filters CNN0, CNN1,..., CNNn, a selector 305f2, and a second stage CNN. A filter 305f3, a qP image generation unit 400f, and a conversion unit 402f are included.

  In the first-stage CNN filter 305f1, the CNN filter CNN1 is a filter optimized for the converted quantization parameter having a value larger than QP1L and smaller than QP1H. Further, the CNN filter CNN2 is a filter optimized for the converted quantization parameter having a value larger than QP2L and smaller than QP2H. Further, the CNN filter CNN3 is a filter optimized for the converted quantization parameter having a value larger than QP3L and smaller than QP3H. Further, the CNN filter CNN4 is a filter optimized for the converted quantization parameter having a value larger than QP4L and smaller than QP4H. Other CNN filters are similar filters.

  As specific examples of thresholds QP1L, QP1H, ... QP4L, QP4H, assign values of QP1L = 0, QP1H = 18, QP2L = 12, QP2H = 30, QP3L = 24, QP3H = 42, QP4L = 36, QP4H = 51 Can do.

  In this case, for example, if the converted quantization parameter is 10, the selector 305f2 selects the CNN filter CNN1. If the converted quantization parameter is 15, the selector 305f2 selects the CNN filter CNN1 and the CNN filter CNN2. If the converted quantization parameter is 20, the selector 305f2 selects the CNN filter CNN2. If the converted quantization parameter is 25, the selector 305f2 selects the CNN filter CNN2 and the CNN filter CNN3. If the converted quantization parameter is 30, the selector 305f2 selects the CNN filter CNN3.

  The second-stage CNN filter 305f3 outputs an input image as a filtered image when the selector 305f2 selects one type of CNN filter, and outputs two input images when the selector 305f2 selects two types of CNN filters. The average value of the image is output as a filtered image.

  Also in this embodiment, the CNN filter in the image decoding apparatus has the same function as the CNN filter 107f in the image encoding apparatus.

  According to the configuration of the fifth embodiment, the part (305f1) that switches the network according to the magnitude of the converted quantization parameter and the part that uses the same network regardless of the magnitude of the converted quantization parameter By using (305f3), the network configuration can be reduced as compared with the configuration in which the entire filter is switched by an encoding parameter such as a quantization parameter. The smaller the network configuration, the smaller the calculation amount and the higher the speed. In addition, the learning parameters are robust, and an appropriate filter process can be performed for many input images. In addition, by overlapping the optimization range of each CNN filter, it is possible to avoid the appearance of visual distortion at the patch boundary when the filter is switched.

<Sixth Embodiment>
The following will describe another embodiment of the present invention with reference to FIG. For convenience of explanation, description of members having the same functions as those described in the embodiment is omitted.

  As described above, the image filter device may use a function for reducing block distortion and a filter for reducing ringing distortion. Also, another filter such as a deblocking filter (DF) that reduces block distortion and a sample adaptive offset (SAO) that reduces ringing distortion may be used in combination.

  In the present embodiment, a configuration in which a deblocking filter (DF) process, a sample adaptive offset (SAO) process, and a CNN filter are used together will be described.

(First example)
FIG. 14A shows a first example of this embodiment. In the first example, the image filter device 305g includes a CNN filter 305g1 and a sample adaptive offset (SAO) 305g2. The CNN filter 305g1 functions as a filter that reduces block distortion.

(Second example)
FIG. 14B shows a second example of this embodiment. In the second example, the image filter device 305h includes a deblocking filter (DF) 305h1 and a CNN filter 305g2. The CNN filter 305h2 functions as a filter that further reduces ringing noise after the deblocking filter.

(Third example)
FIG. 14C shows a third example of this embodiment. In the third example, the image filter device 305i includes a first CNN filter 305i1 and a second CNN filter 305i2. The first CNN filter 305i1 functions as a filter that reduces block distortion, and the second CNN filter 305i2 functions as a filter that further reduces ringing noise after the filter that reduces block distortion.

  In any example, the CNN filter in the image decoding device has the same function as the CNN filter in the image encoding device.

<Seventh Embodiment>
Note that a part of the image encoding device 11 and the image decoding device 31 in the above-described embodiment, for example, the prediction parameter decoding unit 302, the loop filter 305, the predicted image generation unit 308, the inverse quantization / inverse transformation unit 311, and the addition unit 312, prediction image generation unit 101, subtraction unit 102, transform / quantization unit 103, entropy encoding unit 104, inverse quantization / inverse transform unit 105, loop filter 107, encoding parameter determination unit 110, prediction parameter encoding unit 111 may be realized by a computer. In that case, a program for realizing this control function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” is a computer system built in either the image encoding device 11 or the image decoding device 31 and includes hardware such as an OS and peripheral devices. The “computer-readable recording medium” refers to a storage device such as a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a hard disk built in a computer system. Further, the “computer-readable recording medium” is a medium that dynamically holds a program for a short time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line, In such a case, a volatile memory inside a computer system serving as a server or a client may be included and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  Moreover, you may implement | achieve part or all of the image coding apparatus 11 in the embodiment mentioned above, and the image decoding apparatus 31 as integrated circuits, such as LSI (Large Scale Integration). Each functional block of the image encoding device 11 and the image decoding device 31 may be individually made into a processor, or a part or all of them may be integrated into a processor. Further, the method of circuit integration is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. Further, in the case where an integrated circuit technology that replaces LSI appears due to progress in semiconductor technology, an integrated circuit based on the technology may be used.

  As described above, the embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to the above, and various design changes and the like can be made without departing from the scope of the present invention. It is possible to

[Application example]
The image encoding device 11 and the image decoding device 31 described above can be used by being mounted on various devices that perform transmission, reception, recording, and reproduction of moving images. The moving image may be a natural moving image captured by a camera or the like, or an artificial moving image (including CG and GUI) generated by a computer or the like.

  First, it will be described with reference to FIG. 15 that the image encoding device 11 and the image decoding device 31 described above can be used for transmission and reception of moving images.

  FIG. 15A is a block diagram illustrating a configuration of a transmission device PROD_A in which the image encoding device 11 is mounted. As illustrated in (a) of FIG. 15, the transmission device PROD_A modulates a carrier wave with an encoding unit PROD_A1 that obtains encoded data by encoding a moving image, and with the encoded data obtained by the encoding unit PROD_A1 Thus, a modulation unit PROD_A2 that obtains a modulation signal and a transmission unit PROD_A3 that transmits the modulation signal obtained by the modulation unit PROD_A2 are provided. The above-described image encoding device 11 is used as the encoding unit PROD_A1.

  Transmission device PROD_A, as a source of moving images to be input to the encoding unit PROD_A1, a camera PROD_A4 that captures moving images, a recording medium PROD_A5 that records moving images, an input terminal PROD_A6 for inputting moving images from the outside, and An image processing unit A7 that generates or processes an image may be further provided. FIG. 15A illustrates a configuration in which the transmission apparatus PROD_A includes all of these, but some of them may be omitted.

  Note that the recording medium PROD_A5 may be a recording of a non-encoded moving image, or a recording of a moving image encoded by a recording encoding scheme different from the transmission encoding scheme. It may be a thing. In the latter case, a decoding unit (not shown) for decoding the encoded data read from the recording medium PROD_A5 in accordance with the recording encoding method may be interposed between the recording medium PROD_A5 and the encoding unit PROD_A1.

  FIG. 15B is a block diagram illustrating a configuration of the receiving device PROD_B in which the image decoding device 31 is mounted. As shown in FIG. 15B, the receiving device PROD_B includes a receiving unit PROD_B1 that receives a modulated signal, a demodulating unit PROD_B2 that obtains encoded data by demodulating the modulated signal received by the receiving unit PROD_B1, and a demodulator A decoding unit PROD_B3 that obtains a moving image by decoding the encoded data obtained by the unit PROD_B2. The above-described image decoding device 31 is used as the decoding unit PROD_B3.

  The receiving device PROD_B is a display destination PROD_B4 for displaying a moving image, a recording medium PROD_B5 for recording a moving image, and an output terminal for outputting the moving image to the outside as a supply destination of the moving image output by the decoding unit PROD_B3 PROD_B6 may be further provided. FIG. 15B illustrates a configuration in which the receiver PROD_B includes all of these, but some of them may be omitted.

  Note that the recording medium PROD_B5 may be used for recording a non-encoded moving image, or is encoded using a recording encoding method different from the transmission encoding method. May be. In the latter case, an encoding unit (not shown) for encoding the moving image acquired from the decoding unit PROD_B3 according to the recording encoding method may be interposed between the decoding unit PROD_B3 and the recording medium PROD_B5.

  Note that the transmission medium for transmitting the modulation signal may be wireless or wired. Further, the transmission mode for transmitting the modulated signal may be broadcasting (here, a transmission mode in which the transmission destination is not specified in advance) or communication (here, transmission in which the transmission destination is specified in advance). Refers to the embodiment). That is, the transmission of the modulation signal may be realized by any of wireless broadcasting, wired broadcasting, wireless communication, and wired communication.

  For example, a terrestrial digital broadcast broadcasting station (such as broadcasting equipment) / receiving station (such as a television receiver) is an example of a transmitting device PROD_A / receiving device PROD_B that transmits and receives a modulated signal by wireless broadcasting. A broadcasting station (such as broadcasting equipment) / receiving station (such as a television receiver) of cable television broadcasting is an example of a transmitting device PROD_A / receiving device PROD_B that transmits and receives a modulated signal by cable broadcasting.

  Also, a server (workstation, etc.) / Client (television receiver, personal computer, smartphone, etc.) such as a VOD (Video On Demand) service or a video sharing service using the Internet is a transmission device that transmits and receives modulated signals by communication. This is an example of PROD_A / receiving device PROD_B (normally, either a wireless or wired transmission medium is used in a LAN, and a wired transmission medium is used in a WAN). Here, the personal computer includes a desktop PC, a laptop PC, and a tablet PC. The smartphone also includes a multi-function mobile phone terminal.

  Note that the client of the video sharing service has a function of encoding a moving image captured by a camera and uploading it to the server in addition to a function of decoding the encoded data downloaded from the server and displaying it on the display. That is, the client of the video sharing service functions as both the transmission device PROD_A and the reception device PROD_B.

  Next, it will be described with reference to FIG. 16 that the image encoding device 11 and the image decoding device 31 described above can be used for recording and reproduction of moving images.

  FIG. 16A is a block diagram illustrating a configuration of a recording apparatus PROD_C in which the above-described image encoding device 11 is mounted. As shown in FIG. 16 (a), the recording apparatus PROD_C includes an encoding unit PROD_C1 that obtains encoded data by encoding a moving image, and the encoded data obtained by the encoding unit PROD_C1 on a recording medium PROD_M. A writing unit PROD_C2 for writing. The above-described image encoding device 11 is used as the encoding unit PROD_C1.

  The recording medium PROD_M may be of a type built into the recording device PROD_C, such as (1) HDD (Hard Disk Drive) or SSD (Solid State Drive), or (2) SD memory. It may be of the type connected to the recording device PROD_C, such as a card or USB (Universal Serial Bus) flash memory, or (3) DVD (Digital Versatile Disc) or BD (Blu-ray Disc: registration) Or a drive device (not shown) built in the recording device PROD_C.

  In addition, the recording device PROD_C is a camera PROD_C3 that captures moving images as a source of moving images to be input to the encoding unit PROD_C1, an input terminal PROD_C4 for inputting moving images from the outside, and a reception for receiving moving images A unit PROD_C5 and an image processing unit PROD_C6 for generating or processing an image may be further provided. FIG. 16A illustrates a configuration in which the recording apparatus PROD_C includes all of these, but some of them may be omitted.

  The receiving unit PROD_C5 may receive a non-encoded moving image, or may receive encoded data encoded by a transmission encoding scheme different from the recording encoding scheme. You may do. In the latter case, a transmission decoding unit (not shown) that decodes encoded data encoded by the transmission encoding method may be interposed between the reception unit PROD_C5 and the encoding unit PROD_C1.

  Examples of such a recording device PROD_C include a DVD recorder, a BD recorder, an HDD (Hard Disk Drive) recorder, and the like (in this case, the input terminal PROD_C4 or the receiving unit PROD_C5 is a main source of moving images). . In addition, a camcorder (in this case, the camera PROD_C3 is a main source of moving images), a personal computer (in this case, the receiving unit PROD_C5 or the image processing unit C6 is a main source of moving images), a smartphone (this In this case, the camera PROD_C3 or the reception unit PROD_C5 is a main source of moving images), and the like is also an example of such a recording apparatus PROD_C.

  FIG. 16B is a block showing the configuration of the playback device PROD_D in which the above-described image decoding device 31 is mounted. As shown in FIG. 16 (b), the playback device PROD_D reads a moving image by decoding a read unit PROD_D1 that reads encoded data written to the recording medium PROD_M and a read unit PROD_D1 that reads the encoded data. And a decoding unit PROD_D2 to obtain. The above-described image decoding device 31 is used as the decoding unit PROD_D2.

  The recording medium PROD_M may be of the type built into the playback device PROD_D, such as (1) HDD or SSD, or (2) such as an SD memory card or USB flash memory. It may be of the type connected to the playback device PROD_D, or (3) may be loaded into a drive device (not shown) built in the playback device PROD_D, such as a DVD or BD. Good.

  In addition, the playback device PROD_D has a display unit PROD_D3 that displays a moving image as a supply destination of the moving image output by the decoding unit PROD_D2, an output terminal PROD_D4 that outputs the moving image to the outside, and a transmission unit that transmits the moving image. PROD_D5 may be further provided. FIG. 16B illustrates a configuration in which the playback apparatus PROD_D includes all of these, but some of the configurations may be omitted.

  The transmission unit PROD_D5 may transmit a non-encoded moving image, or transmits encoded data encoded by a transmission encoding scheme different from the recording encoding scheme. You may do. In the latter case, it is preferable to interpose an encoding unit (not shown) that encodes a moving image using a transmission encoding method between the decoding unit PROD_D2 and the transmission unit PROD_D5.

  Examples of such a playback device PROD_D include a DVD player, a BD player, and an HDD player (in this case, an output terminal PROD_D4 to which a television receiver or the like is connected is a main moving image supply destination). . In addition, a television receiver (in this case, the display PROD_D3 is a main supply destination of moving images), a digital signage (also referred to as an electronic signboard or an electronic bulletin board), and the display PROD_D3 or the transmission unit PROD_D5 is the main supply of moving images. Desktop PC (in this case, output terminal PROD_D4 or transmission unit PROD_D5 is the main video source), laptop or tablet PC (in this case, display PROD_D3 or transmission unit PROD_D5 is video) A smartphone (which is a main image supply destination), a smartphone (in this case, the display PROD_D3 or the transmission unit PROD_D5 is a main moving image supply destination), and the like are also examples of such a playback device PROD_D.

(Hardware implementation and software implementation)
Each block of the image decoding device 31 and the image encoding device 11 described above may be realized in hardware by a logic circuit formed on an integrated circuit (IC chip), or may be a CPU (Central Processing Unit). You may implement | achieve by software using.

  In the latter case, each device includes a CPU that executes instructions of a program that realizes each function, a ROM (Read Only Memory) that stores the program, a RAM (Random Access Memory) that expands the program, the program, and various types A storage device (recording medium) such as a memory for storing data is provided. The object of the embodiment of the present invention is a record in which the program code (execution format program, intermediate code program, source program) of the control program for each of the above devices, which is software that realizes the functions described above, is recorded so as to be readable by a computer. This can also be achieved by supplying a medium to each of the above devices, and reading and executing the program code recorded on the recording medium by the computer (or CPU or MPU).

  Examples of the recording medium include magnetic tapes and cassette tapes, magnetic disks such as floppy (registered trademark) disks / hard disks, CD-ROM (Compact Disc Read-Only Memory) / MO disks (Magneto-Optical discs), and the like. ) / MD (Mini Disc) / DVD (Digital Versatile Disc) / CD-R (CD Recordable) / Blu-ray Disc (registered trademark) and other optical discs, IC cards (including memory cards) / Cards such as optical cards, mask ROM / EPROM (Erasable Programmable Read-Only Memory) / EEPROM (Electrically Erasable and Programmable Read-Only Memory: registered trademark) / Semiconductor memories such as flash ROM, or PLD (Programmable logic device) ) Or FPGA (Field Programmable Gate Array).

  Further, each of the above devices may be configured to be connectable to a communication network, and the program code may be supplied via the communication network. The communication network is not particularly limited as long as it can transmit the program code. For example, the Internet, Intranet, Extranet, LAN (Local Area Network), ISDN (Integrated Services Digital Network), VAN (Value-Added Network), CATV (Community Antenna television / Cable Television) communication network, Virtual Private Network (Virtual Private Network) Network), telephone line network, mobile communication network, satellite communication network, and the like. The transmission medium constituting the communication network may be any medium that can transmit the program code, and is not limited to a specific configuration or type. For example, IEEE (Institute of Electrical and Electronic Engineers) 1394, USB, power line carrier, cable TV line, telephone line, ADSL (Asymmetric Digital Subscriber Line) line, etc. wired, such as IrDA (Infrared Data Association) or remote control , BlueTooth (registered trademark), IEEE802.11 wireless, HDR (High Data Rate), NFC (Near Field Communication), DLNA (Digital Living Network Alliance: registered trademark), mobile phone network, satellite line, terrestrial digital broadcasting network, etc. It can also be used wirelessly. The embodiment of the present invention can also be realized in the form of a computer data signal embedded in a carrier wave in which the program code is embodied by electronic transmission.

  The embodiments of the present invention are not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. That is, embodiments obtained by combining technical means appropriately modified within the scope of the claims are also included in the technical scope of the present invention.

[Summary]
In the image filter device (CNN filters 107 and 305) according to the first aspect of the present invention, at least one or more types of first type input image data having a luminance or color difference as a pixel value are input, and the luminance or color difference is set as a pixel value. A neural network unit that outputs one or a plurality of first type output image data, and one or a plurality of second type input image data having a quantization parameter as a pixel value, and the second type input image A conversion unit (402) that generates a second type of input image data after conversion by nonlinearly converting the data, and the neural network unit (404) has the second type of input after the conversion. Image data is also input.

  The conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 2 of the present invention is the inverse quantization or quantization when generating the first type of input image data in the above aspect 1. An image having a pixel value as the quantization step used in is generated as the second type of input image data after conversion.

  The conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 3 of the present invention is the inverse quantization or quantization when generating the first type of input image data in the above aspect 1. An image having a pixel value that is the inverse of the quantization step used in is generated as the second type of input image data after conversion.

  In the aspect 1, the conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 4 of the present invention performs the following on each quantization parameter included in the second type of input image data. By performing any one of logarithmic transformation, exponential transformation, and reciprocal transformation, the second type of input image data after the transformation is generated.

  The conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 5 of the present invention includes a polynomial for each quantization parameter included in the second type of input image data in aspect 1 above. The second type of input image data after the conversion is generated by performing non-linear conversion using.

  The conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 6 of the present invention standardizes the quantization parameter that is the pixel value of the second type of input image data in aspect 1 above. By performing the conversion, the second type of input image data after the conversion is generated.

  In the image filter device (CNN filters 107 and 305) according to the seventh aspect of the present invention, at least one or a plurality of first type input image data having a pixel value of luminance or color difference is input, and luminance or pigment is used as the pixel value. A neural network unit (404) that outputs one or a plurality of first-type output image data, and one or a plurality of second-type input image data having a quantization parameter as a pixel value. A conversion unit (402) that generates the second type of input image data after conversion by converting the second type of input image data with reference to an encoding parameter for generating an image. The second input image data after the conversion is also input to the neural network unit.

  The transform unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 8 of the present invention is configured such that intra prediction is applied to the target block by referring to the encoding parameter in aspect 7 above. If the target block is determined to be a block to which intra prediction is applied, the transform unit calculates quantization parameters of all pixels included in the target block. By converting, the second type of input image data after the conversion is generated.

  In the aspect 7, the conversion unit (402) included in the image filter device (CNN filters 107 and 305) according to aspect 9 of the present invention specifies a block boundary by referring to the encoding parameter, and specifies the specified block The second input image data after the conversion is generated by changing the quantization parameter of the pixel adjacent to the boundary.

  An image decoding device (31) according to an aspect 10 of the present invention is an image decoding device that decodes an image, and the image filter device according to any one of the above aspects 1 to 9 is used as a filter that acts on a decoded image. Prepare.

  An image encoding device (11) according to an aspect 11 of the present invention is an image encoding device that encodes an image, and is described in any one of the above aspects 1 to 9 as a filter that acts on a locally decoded image. An image filter device is provided.

  The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope shown in the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments. Is also included in the technical scope of the present invention. Furthermore, a new technical feature can be formed by combining the technical means disclosed in each embodiment.

  Embodiments of the present invention can be preferably applied to an image decoding apparatus that decodes encoded data in which image data is encoded, and an image encoding apparatus that generates encoded data in which image data is encoded. it can. Further, the present invention can be suitably applied to the data structure of encoded data generated by an image encoding device and referenced by the image decoding device.

DESCRIPTION OF SYMBOLS 11 Image coding apparatus 31 Image decoding apparatus 402 Conversion part 404 Neural network part 107 CNN filter (image filter apparatus)

Claims (11)

In the image filter device,
A neural network unit that receives at least one or a plurality of first type input image data having a luminance or color difference as a pixel value and outputs one or a plurality of first type output image data having a luminance or a color difference as a pixel value; ,
One or a plurality of second type input image data having a quantization parameter as a pixel value is input, and the second type input image data is nonlinearly converted to generate a second type input image data after conversion. A conversion unit to
With
2. The image filter device according to claim 1, wherein the second type of input image data after the conversion is input to the neural network unit.   The conversion unit uses, as the second type of input image data after the conversion, an image having a pixel value corresponding to a quantization step used in inverse quantization or quantization when generating the first type of input image data. The image filter device according to claim 1, wherein the image filter device is generated.   The conversion unit performs an inverse quantization when generating the first type of input image data, or an image having a pixel value as an inverse number of a quantization step used in the quantization, and a second type of input image after the conversion. The image filter device according to claim 1, wherein the image filter device is generated as data.   The conversion unit performs any one of logarithmic conversion, exponential conversion, and reciprocal conversion on each quantization parameter included in the second type of input image data, thereby performing the second type after the conversion. The image filter apparatus according to claim 1, wherein the input image data is generated.   The conversion unit generates the second type of input image data after the conversion by performing nonlinear conversion using a polynomial for each quantization parameter included in the second type of input image data. The image filter device according to claim 1, wherein:   The conversion unit generates the second type of input image data after the conversion by performing a conversion that standardizes a quantization parameter that is a pixel value of the second type of input image data. Item 2. The image filter device according to Item 1. In the image filter device,
A neural network unit that receives at least one or a plurality of first type input image data having a luminance or color difference as a pixel value and outputs one or a plurality of first type output image data having a luminance or a pigment as a pixel value; ,
One or a plurality of second type input image data having a quantization parameter as a pixel value is input, and the second type input image data is obtained by referring to an encoding parameter for generating a predicted image or a difference image. A conversion unit that generates the second type of input image data after conversion by converting;
With
2. The image filter device according to claim 1, wherein the second input image data after the conversion is also input to the neural network unit. The conversion unit is
By referring to the encoding parameter, it is determined whether the target block is a block to which intra prediction is applied,
When it is determined that the target block is a block to which intra prediction is applied,
The conversion unit converts the quantization parameters of all pixels included in the target block,
8. The image filter device according to claim 7, wherein the second type of input image data after the conversion is generated. The conversion unit is
A block boundary is specified by referring to the encoding parameter, and the converted second input image data is generated by changing a quantization parameter of a pixel adjacent to the specified block boundary. The image filter device according to claim 7.   An image decoding apparatus that decodes an image, comprising the image filter apparatus according to claim 1 as a filter that acts on a decoded image.   An image encoding apparatus that encodes an image, comprising the image filter apparatus according to claim 1 as a filter that acts on a locally decoded image.

Images