CROSS-REFERENCE TO RELATED APPLICATION

[0001] The present application claims priority to U.S. Provisional Application No. 63/127,987, entitled “Neural Network Based Image Filtering for Video Coding,” filed on Dec. 18, 2020, the entirety of which is incorporated by reference for all purposes.

FIELD

[0002] The present disclosure relates to video coding, and in particular but not limited to, methods and apparatus on video coding using neural network based model filtering.

BACKGROUND

[0003] Various video coding techniques may be used to compress video data. Video coding is performed according to one or more video coding standards. For example, video coding standards include versatile video coding (VVC), joint exploration test model (JEM), high-efficiency video coding (H.265/HEVC), advanced video coding (H.264/AVC), moving picture expert group (MPEG) coding, or the like. Video coding generally utilizes prediction methods (e.g., interprediction, intra-prediction, or the like) that take advantage of redundancy present in video images or sequences. An important goal of video coding techniques is to compress video data into a form that uses a lower bit rate, while avoiding or minimizing degradations to video quality.

[0004] The first version of the HEVC standard was finalized in October 2013, which offers approximately 50% bit-rate saving or equivalent perceptual quality compared to the prior generation video coding standard H.264/MPEG AVC. Although the HEVC standard provides significant coding improvements than its predecessor, there is evidence that superior coding efficiency can be achieved with additional coding tools over HEVC. Based on that, both VCEG and MPEG started the exploration work of new coding technologies for future video coding standardization, one Joint Video Exploration Team (JVET) was formed in Oct. 2015 by ITU-T VECG and ISO/IEC MPEG to begin significant study of advanced technologies that could enable substantial enhancement of coding efficiency. One reference software called joint exploration model (JEM) was maintained by the JVET by integrating several additional coding tools on top of the HEVC test model (HM). [0005] The joint call for proposals (CfP) on video compression with capability beyond HEVC was issued by ITU-T and ISO/IEC. 23 CfP responses were received and evaluated at the 10-th JVET meeting, which demonstrated compression efficiency gain over the HEVC around 40%. Based on such evaluation results, the JVET launched a new project to develop the new generation video coding standard that is named as Versatile Video Coding (VVC). One reference software codebase, called VVC test model (VTM), was established for demonstrating a reference implementation of the VVC standard.

SUMMARY

[0006] The present disclosure provides examples of techniques relating to improving the video coding efficiency by using neural network based model filtering.

[0007] According to a first aspect of the present disclosure, there is provided a method for image filtering in video coding using a neural network. The method includes: loading, a plurality of quantization parameter (QP) map (QpMap) values at a plurality of QpMap channels into the neural network, and adjusting, according to a QP scaling factor, the plurality of QpMap values for the neural network to learn and filter a current input frame to the neural network.

[0008] According to a second aspect of the present disclosure, there is provided a method for adjusting outputs of a neural network based image filter in video coding. The method includes: adaptively selecting, by an encoder, an adjustment factor from a plurality of values, where the plurality of values are coded in a picture header or a sequence parameter set associate a current input frame to be coded; and adjusting, according to the adjustment factor, an output of the neural network based image filter.

[0009] According to a third aspect of the present disclosure, there is provided a method for training a neural network based image filter. The method includes: training the neural network based image filter according to a plurality of intra-coded frames to obtain a first model and a first data set encoded with the plurality of intra-coded frames; generating a plurality of I/B/P frames including one or more intra-coded frames, one or more bi-directional predicted frames, and one or more predicted frames; obtaining a second data set by encoding the plurality of I/B/P pictures using the first model; and obtaining a second model based on the second data set.

[0010] According to a fourth aspect of the present disclosure, there is provided an apparatus for image filtering in video coding using a neural network. The apparatus includes one or more processors and a memory configured to store instructions executable by the one or more processors. Further, the one or more processors, upon execution of the instructions, are configured to perform the method according to the first aspect.

[0011] According to a fifth aspect of the present disclosure, there is provided an apparatus for adjusting outputs of a neural network based image filter in video coding. The apparatus includes one or more processors and a memory configured to store instructions executable by the one or more processors. Further, the one or more processors, upon execution of the instructions, are configured to perform the method according to the second aspect.

[0012] According to a sixth aspect of the present disclosure, there is provided an apparatus for training a neural network based image filter. The apparatus includes one or more processors and a memory configured to store instructions executable by the one or more processors. Further, the one or more processors, upon execution of the instructions, are configured to perform the method according to the third aspect.

[0013] According to a seventh aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer-executable instructions that, when executed by one or more computer processors, causing the one or more computer processors to perform the method according to the first aspect.

[0014] According to an eighth aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer-executable instructions that, when executed by one or more computer processors, causing the one or more computer processors to perform the method according to the second aspect.

[0015] According to a ninth aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer-executable instructions that, when executed by one or more computer processors, causing the one or more computer processors to perform the method according to the third aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

[0016] A more particular description of the examples of the present disclosure will be rendered by reference to specific examples illustrated in the appended drawings. Given that these drawings depict only some examples and are not therefore considered to be limiting in scope, the examples will be described and explained with additional specificity and details through the use of the accompanying drawings.

[0017] FIG. l is a block diagram illustrating a block-based video encoder in accordance with some implementations of the present disclosure.

[0018] FIG. 2 is a block diagram illustrating a block-based video decoder in accordance with some implementations of the present disclosure.

[0019] FIG. 3A is schematic diagram illustrating quaternary partitioning tree splitting mode in accordance with some implementations of the present disclosure.

[0020] FIG. 3B is schematic diagram illustrating vertical binary partitioning tree splitting mode in accordance with some implementations of the present disclosure.

[0021] FIG. 3C is schematic diagram illustrating horizontal binary partitioning tree splitting mode in accordance with some implementations of the present disclosure.

[0022] FIG. 3D is schematic diagram illustrating vertical ternary partitioning tree splitting mode in accordance with some implementations of the present disclosure.

[0023] FIG. 3E is schematic diagram illustrating horizontal ternary partitioning tree splitting mode in accordance with some implementations of the present disclosure.

[0024] FIG. 4 illustrates a simple FC-NN consisting of input layer, output layer, and multiple hidden layers in accordance with some implementations of the present disclosure.

[0025] FIG. 5A illustrates an FC-NN with two hidden layers in accordance with some implementations of the present disclosure.

[0026] FIG. 5B illustrates an example of CNN in which the dimension of the second hidden layer is [W, H, Depth] in accordance with some implementations of the present disclosure.

[0027] FIG. 6 illustrates an example of applying spatial filters with an input image in accordance with some implementations of the present disclosure.

[0028] FIG. 7A illustrates a single image super-resolution (ResNet) including a residual block as the element of ResNet that is elementwise added with its input by identity connection in accordance with some implementations of the present disclosure.

[0029] FIG. 7B illustrates an example of ResNet by staking residual modules in accordance with some implementations of the present disclosure.

[0030] FIG. 8A illustrates an example of ResNet including a plurality of residual blocks with global identity connection in accordance with some implementations in the present disclosure. [0031] FIG. 8B illustrates another example of ResNet stacks multiple residual blocks in order to further improve the video coding efficiency in accordance with some implementations in the present disclosure.

[0032] FIG. 8C illustrates another example of ResNet tackling single-image super-resolution (SISR) by aggregating the output of residual blocks in accordance with some implementations in the present disclosure.

[0033] FIG. 9 illustrates a typical neural network based model to perform image filtering for video coding in accordance with some implementations in the present disclosure.

[0034] FIG. 10 illustrates region-based feature map resolution control in accordance with some implementations in the present disclosure.

[0035] FIG. 11 illustrates a typical QP -independent neural network model in accordance with some implementations of the present disclosure.

[0036] FIG. 12A illustrates an example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure.

[0037] FIG. 12B illustrates an example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure.

[0038] FIG. 13 illustrates Chroma up-sampling in region 2 of a neural network in accordance with some implementations of the present disclosure.

[0039] FIG. 14 illustrates Luma down-sampling in region 2 of a neural network in accordance with some implementations of the present disclosure.

[0040] FIG. 15 illustrates Luma down-sampling in region 1 of a neural network in accordance with some implementations of the present disclosure.

[0041] FIG. 16A illustrates another example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure.

[0042] FIG. 16B illustrates another example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure.

[0043] FIG. 17 is a block diagram illustrating an apparatus for image filtering in video coding using a neural network in accordance with some implementations of the present disclosure.

[0044] FIG. 18 is a flowchart illustrating a process for image filtering in video coding using a neural network in accordance with some implementations of the present disclosure. [0045] FIG. 19 is a flowchart illustrating an example process for adjusting outputs of a neural network based image filter in video coding in accordance with the present disclosure.

[0046] FIG. 20 is a flowchart illustrating an example process for training a neural network based image filter in accordance with the present disclosure.

DETAILED DESCRIPTION

[0047] Reference will now be made in detail to specific implementations, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous nonlimiting specific details are set forth in order to assist in understanding the subject matter presented herein. But it will be apparent to one of ordinary skill in the art that various alternatives may be used. For example, it will be apparent to one of ordinary skill in the art that the subject matter presented herein can be implemented on many types of electronic devices with digital video capabilities.

[0048] Reference throughout this specification to “one embodiment,” “an embodiment,” “an example,” “some embodiments,” “some examples,” or similar language means that a particular feature, structure, or characteristic described is included in at least one embodiment or example. Features, structures, elements, or characteristics described in connection with one or some embodiments are also applicable to other embodiments, unless expressly specified otherwise.

[0049] Throughout the disclosure, the terms “first,” “second,” “third,” etc. are all used as nomenclature only for references to relevant elements, e.g., devices, components, compositions, steps, etc., without implying any spatial or chronological orders, unless expressly specified otherwise. For example, a “first device” and a “second device” may refer to two separately formed devices, or two parts, components, or operational states of a same device, and may be named arbitrarily.

[0050] The terms “module,” “sub-module,” “circuit,” “sub-circuit,” “circuitry,” “sub-circuitry,” “unit,” or “sub-unit” may include memory (shared, dedicated, or group) that stores code or instructions that can be executed by one or more processors. A module may include one or more circuits with or without stored code or instructions. The module or circuit may include one or more components that are directly or indirectly connected. These components may or may not be physically attached to, or located adjacent to, one another. [0051] As used herein, the term “if’ or “when” may be understood to mean “upon” or “in response to” depending on the context. These terms, if appear in a claim, may not indicate that the relevant limitations or features are conditional or optional. For example, a method may comprise steps of i) when or if condition X is present, function or action X’ is performed, and ii) when or if condition Y is present, function or action Y’ is performed. The method may be implemented with both the capability of performing function or action X’, and the capability of performing function or action Y’ . Thus, the functions X’ and Y’ may both be performed, at different times, on multiple executions of the method.

[0052] A unit or module may be implemented purely by software, purely by hardware, or by a combination of hardware and software. In a pure software implementation, for example, the unit or module may include functionally related code blocks or software components, that are directly or indirectly linked together, so as to perform a particular function.

[0053] Like HEVC, VVC is built upon the block-based hybrid video coding framework. FIG. 1 is a block diagram illustrating a block-based video encoder in accordance with some implementations of the present disclosure. In the encoder 100, the input video signal is processed block by block, called coding units (CUs). In VTM-1.0, a CU can be up to 128x128 pixels. However, different from the HEVC which partitions blocks only based on quad-trees, in VVC, one coding tree unit (CTU) is split into CUs to adapt to varying local characteristics based on quad/binary/ternary-tree. Additionally, the concept of multiple partition unit type in the HEVC is removed, i.e., the separation of CU, prediction unit (PU) and transform unit (TU) does not exist in the VVC anymore; instead, each CU is always used as the basic unit for both prediction and transform without further partitions. In the multi-type tree structure, one CTU is firstly partitioned by a quad-tree structure. Then, each quad-tree leaf node can be further partitioned by a binary and ternary tree structure.

[0054] FIGS. 3A-3E are schematic diagrams illustrating multi-type tree splitting modes in accordance with some implementations of the present disclosure. FIGS. 3A-3E respectively show five splitting types including quaternary partitioning (FIG. 3 A), vertical binary partitioning (FIG. 3B), horizontal binary partitioning (FIG. 3C), vertical ternary partitioning (FIG. 3D), and horizontal ternary partitioning (FIG. 3E).

[0055] For each given video block, spatial prediction and/or temporal prediction may be performed. Spatial prediction (or “intra prediction”) uses pixels from the samples of already coded neighboring blocks (which are called reference samples) in the same video picture/ slice to predict the current video block. Spatial prediction reduces spatial redundancy inherent in the video signal. Temporal prediction (also referred to as “inter prediction” or “motion compensated prediction”) uses reconstructed pixels from the already coded video pictures to predict the current video block. Temporal prediction reduces temporal redundancy inherent in the video signal. Temporal prediction signal for a given CU is usually signaled by one or more motion vectors (MVs) which indicate the amount and the direction of motion between the current CU and its temporal reference. Also, if multiple reference pictures are supported, one reference picture index is additionally sent, which is used to identify from which reference picture in the reference picture store the temporal prediction signal comes.

[0056] After spatial and/or temporal prediction, an intra/inter mode decision circuitry 121 in the encoder 100 chooses the best prediction mode, for example based on the rate-distortion optimization method. The block predictor 120 is then subtracted from the current video block; and the resulting prediction residual is de-correlated using the transform circuitry 102 and the quantization circuitry 104. The resulting quantized residual coefficients are inverse quantized by the inverse quantization circuitry 116 and inverse transformed by the inverse transform circuitry 118 to form the reconstructed residual, which is then added back to the prediction block to form the reconstructed signal of the CU. Further, in-loop filtering 115, such as a deblocking filter, a sample adaptive offset (SAO), and/or an adaptive in-loop filter (ALF) may be applied on the reconstructed CU before it is put in the reference picture store of the picture buffer 117 and used to code future video blocks. To form the output video bitstream 114, coding mode (inter or intra), prediction mode information, motion information, and quantized residual coefficients are all sent to the entropy coding unit 106 to be further compressed and packed to form the bit-stream.

[0057] For example, a deblocking filter is available in AVC, HEVC as well as the now-current version of VVC. In HEVC, an additional in-loop filter called SAO is defined to further improve coding efficiency. In the now-current version of the VVC standard, yet another in-loop filter called ALF is being actively investigated, and it has a good chance of being included in the final standard. [0058] These in-loop filter operations are optional. Performing these operations helps to improve coding efficiency and visual quality. They may also be turned off as a decision rendered by the encoder 100 to save computational complexity. [0059] It should be noted that intra prediction is usually based on unfiltered reconstructed pixels, while inter prediction is based on filtered reconstructed pixels if these filter options are turned on by the encoder 100.

[0060] FIG. 2 is a block diagram illustrating a block-based video decoder 200 which may be used in conjunction with many video coding standards. This decoder 200 is similar to the reconstruction-related section residing in the encoder 100 of FIG. 1. In the decoder 200, an incoming video bitstream 201 is first decoded through an Entropy Decoding 202 to derive quantized coefficient levels and prediction-related information. The quantized coefficient levels are then processed through an Inverse Quantization 204 and an Inverse Transform 206 to obtain a reconstructed prediction residual. A block predictor mechanism, implemented in an Intra/inter Mode Selector 212, is configured to perform either an Intra Prediction 208, or a Motion Compensation 210, based on decoded prediction information. A set of unfiltered reconstructed pixels are obtained by summing up the reconstructed prediction residual from the Inverse Transform 206 and a predictive output generated by the block predictor mechanism, using a summer 214.

[0061] The reconstructed block may further go through an In-Loop Filter 209 before it is stored in a Picture Buffer 213 which functions as a reference picture store. The reconstructed video in the Picture Buffer 213 may be sent to drive a display device, as well as used to predict future video blocks. In situations where the In-Loop Filter 209 is turned on, a filtering operation is performed on these reconstructed pixels to derive a final reconstructed Video Output 222.

[0062] The present disclosure is to improve the image filtering design of the above-mentioned video coding standards or techniques. The proposed filtering method in the present disclosure is neural network based, which may be applied as part of the in-loop filtering, e.g., between the deblocking filter and sample adaptive offset (SAO), or as part of post-loop filtering to improve the current video coding techniques, or as part of post-processing filtering after the current video coding techniques.

[0063] The neural network techniques, e.g., fully connected neural network (FC-NN), convolutional neural network (CNN), and long short-term memory network (LSTM), have already achieved significant success in many research domains, including computer vision and video understanding. Fully-connected Neural Network (FC-NN)

[0064] FIG. 4 illustrates a simple FC-NN consisting of input layer, output layer, and multiple hidden layers in accordance with some implementations of the present disclosure. At k-th layer, the output f ^k x ^k ~ ¹ , W ^k , B ^k ), is generated by where x ^k-1 G R ^M is the output of (k-l)-th layer, W ^k E R ^M * ^N and B ^k E R ^N are the weight and the bias at k-th layer. <?(■) is the activation function, e.g., the Rectified Linear Unit (ReLU) function as defined in Eq. (3).

Therefore, the general form of a K-layer FC-NN is written as

[0065] According to the universal approximation hypothesizes and Eq. (4), given any continuous function g(x) and some E > 0, there exists a neural network (x) with a reasonable choice of non-linearity e.g., ReLU, such that Vx, |^(x) — (x)| < E. Therefore, many empirical studies applied neural network as an approximator to mimic a model with hidden variables in order to extract explainable features under the surfaces. For example, applying in image recognition, FC- NN helps researchers to construct a system that understands not just a single pixel, but increasingly much deeper and complex sub-structures, e.g., edges, textures, geometric shapes, and objects. Convolutional Neural Network (CNN)

[0066] FIG. 5A illustrates an FC-NN with two hidden layers in accordance with some implementations of the present disclosure. CNN, a popular neural network architecture for image or video applications, is very similar to the FC-NN as shown in FIG. 5A, which includes weights and bias metrices. A CNN can be seen as a 3-D version of neural network. FIG. 5B illustrates an example of CNN in which the dimension of the second hidden layer is [W, H, Depth] in accordance with some implementations of the present disclosure. In FIG. 5B, neurons are arranged in 3- Dimensional structure (width, height, and depth) to form a CNN, and the second hidden layer is visualized. In this example, the input layer holds input image or video frames therefore its width and height are same as input data. To apply with image or video applications, each neuron in CNN is a spatial filter element with extended depth aligned with its input, e.g., the depth is 3 if there are 3 color components in input images.

[0067] FIG. 6 illustrates an example of applying spatial filters with an input image in accordance with some implementations of the present disclosure. As shown in FIG. 6, the dimension of basic element in CNN is defined as [Filter _width , Filter _height , Input _depth , Output _depth ] and set to [5, 5, 3, 4] in this example. Each spatial filter performs 2-dimensional spatial convolution with 5*5*3 weights on an input image. The input image may be a 64x64x3 image. Then, 4 convolutional results are outputted. Therefore, the dimension of filtered results is [64+4, 64+4, 4] if padding the boundary with additional 2 pixels.

Residual Network (ResNet)

[0068] In image classification, the accuracy is saturated and degrades rapidly when the depth of neural network increases. To be more specifically, adding more layers on deep neural network results in higher training error because the gradient is gradually vanishing along the deep network and toward to zero gradient at the end. Then, the ResNet composed of residual blocks comes to resolve the degradation problem by introducing the identity connection.

[0069] FIG. 7A illustrates a ResNet including a residual block as the element of ResNet that is elementwise added with its input by identity connection in accordance with some implementations of the present disclosure. As shown in FIG. 7A, a basic module of ResNet is consist of the residual block and the identity connection. According to the universal approximation hypothesizes, given an input x, weighted layers with activation function in residual block approximate a hidden function F(x) rather than the output H (x) = F(x) + x.

[0070] By stacking non-linear multi-layer neural network, the residual block explores the features that represent the local characteristic of input images. Without introducing neither additional parameters and computational complexity, the identity connection is proven to make deep learning network trainable by skip one or more non-linear weighted layers as shown in FIG. 7A. Skipping weighted layers, the differential output of the residual layers can be written as [0071] Therefore, even if the differential term is gradually decreasing toward zero, the identity term can still carry on and pass the input to next layer instead of stuck at zero gradient as well as blocking information propagation. If a neuron cannot propagate information to next neuron, it is seen as dead neuron, which is non-trainable element in neural network. After addition, another non-linear activation function can be applied as well. FIG. 7B illustrates an example of ResNet by staking residual modules in accordance with some implementations of the present disclosure. As shown in FIG. 7B, the residual features are fused with the identity features before propagating to the next module.

Variations of ResNet

[0072] In FIGS. 8A-8B, several variations of ResNet were proposed to improve the recovered image quality for single image super-resolution (SISR) and increase the accuracy for image classification. FIG. 8A illustrates an example of ResNet including a plurality of residual blocks with global identity connection in accordance with some implementations in the present disclosure. In FIG. 8A, a variation of ResNet is proposed to enhance the visual quality of the up-sampled images. Specifically, a global identity connection is applied from the input of first residual block to the output of last residual block in order to facilitate the convergence of training procedure.

[0073] FIG. 8B illustrates another example of ResNet stacks multiple residual blocks in order to further improve the video coding efficiency in accordance with some implementations in the present disclosure. Each residual block directly propagates its own input to the following unit by concatenation operation. In other words, each intermediate block can receive multi-hierarchical information from its preceding units because multi-level information can flow through the identical connection. The parameter of each residual block in FIG. 8B is linearly increased with the number of layers because the concatenation operation.

[0074] In FIGS. 8A-8B, before the residual information can be propagated to later module, the residual features have to go through one or several modules. Due to identity connection, these residual features can be rapidly coupled with identity features at certain layer and stop propagating to succeeding module. Therefore, the residual features in previous two variations are limited locally and results in performance degradation.

[0075] FIG. 8C illustrates another example of ResNet tackling single-image super-resolution (SISR) by aggregating the output of residual blocks in accordance with some implementations in the present disclosure. In FIG. 8C, the output of the last residual block is concatenated with all the output of previous three modules. Before applied with element-wise addition with the input of first residual block, the concatenated hierarchical features are fused by convolutional operation. Different from the first two variations, aggregated ResNet make non-local features applicable to the last residual module so that the hierarchical information can be propagated to succeeding blocks, achieving the feature representation in a more discriminative way.

[0076] In the present disclosure, methods and apparatus related to neural network based image filtering are proposed to further improve the coding efficiency of current hybrid video coding. The proposed methods and apparatus may be applied as part of the in-loop filtering, e.g., between the deblocking filter and sample adaptive offset (SAO) as shown in FIG. 2 or as post-loop filtering to improve the current video coding techniques, or as post-processing filtering after the current video coding techniques.

[0077] FIG. 9 illustrates a typical neural network based model to perform image filtering for video coding in accordance with some implementations in the present disclosure. The YUV components may be provided to the neural network model in parallel. This paralleled input of YUV components may be beneficial not only for reducing processing delay but also for the neural network model to learn the correlations among collocated YUV information, e.g., cross-components filtering and/or luma guided chroma filtering. The on/off control of this neural network model based filter may be performed at CTU level for reasonable trade-off between control granularity and signaling overhead.

Feature map resolution alignment

[0078] When the CTU level YUV information is provided to the neural network model filter as shown in FIG. 9, the resolutions of the YUV CTU patches may or may not be the same. For example, if the encoded video content is YUV420, the resolutions of the three collocated YUV patches may not be the same. In this case, resolution alignment is needed. For easier illustration purpose, all the proposed methods and devices in this disclosure are assuming the video content is YUV420. For different content format, e.g., YUV422, YUV444, the proposed methods may be easily extended.

[0079] In some examples, the resolution alignment may be performed before the YUV patches enter the neural network.

[0080] In some examples, one 128x128 Y patch may be down-sampled into one 64x64 patch or four 64x64 patches. When four 64x64 patches are generated, all the information of the original 128x128 patch may be kept and distributed in the four patches. The method used for the information distribution of the original 128x128 patch may be partition based, e.g., one 64x64 patch may be from the top-left of the original 128x128 patch, and another 64x64 patch may be from the top-right of the original 128x128 patch. Alternatively, the method used for the information distribution of the original 128x128 patch may be interleave based, e.g., every four adjacent samples of the original 128x128 patch are evenly distributed in the four 64x64 patches. [0081] In some examples, one 64x64 U or V patch may be up-sampled into one 128x128 patch. [0082] In some examples, the resolution alignment may be performed after the YUV patches enter the neural network. In one example, the Y input resolution may be decreased to match the UV input. One way to achieve this is to use convolution layers with doubled stride size compared to UV input. In this example, at the end of the neural network, a resolution increase layer is needed to scale up the Y content such that the output of the model has the same resolution as the input. One way to achieve this is to use pixel shuffle layer to scale up the Y resolution. In another example, the UV input resolution may be increased to match the Y input. One way to achieve is to use pixel shuffle layer to scale up UV at the beginning of the neural network, and then scale down at the end of the neural network.

Feature map resolution control

[0083] Feature map resolution proportionally affects the neural network processing overhead, but may not be proportionally affect the performance of the neural network. In order to control the computation complexity of the model filtering, different solutions may be available, e.g., number of residual blocks, number of input and output channels of the convolution layers at each residual block. Resolution control of the feature map in a convolution layer is another effective option to control the computation complexity.

[0084] FIG. 10 illustrates region-based feature map resolution control in accordance with some implementations in the present disclosure. As shown in FIG. 10, three regions may be used to adjust feature map resolution for computation complexity control. In the region 1, resolution of the input YUV patches are determined and corresponding scale up/down operations are performed. For example, the up/down sampling methods introduced in “Feature map resolution alignment.” An example is shown in FIG. 15.

[0085] FIG. 15 illustrates Luma down-sampling in region 1 of a neural network in accordance with some implementations of the present disclosure. As shown in FIG. 15, the original Y patch is down-sampled into four down-sampled Y patches in region 1 before entering the neural network. For example, one 128x128 Y patch may be down-sampled into four 64x64 Y patches. The inverse operation, e.g., up-sample, is performed in region 1 after the processing of the neural network finishes. As shown in FIG. 15, the four down-sampled Y patches outputted by the neural network are up-sampled into one original Y patch. For example, the four 64x64 Y patches may be up- sampled into one 128x128 Y patch.

[0086] In region 2, resolution of the input YUV patches are determined and corresponding scale up/down operations are performed before YUV concatenation. As this region is located at the beginning of the neural network, if scale down operations are performed, input information may be lost significantly and overall performance after model training may be compromised. Two examples are respectively shown in FIGS. 13-14.

[0087] FIG. 13 illustrates Chroma up-sampling in region 2 of a neural network in accordance with some implementations of the present disclosure. As shown in FIG. 13, the UV patches, after entering the neural network, are scaled up by a corresponding convolution block or layer in region 2 of the neural network. An inverse operation, e.g., scaling-down, is performed on the corresponding UV patches outputted by the last residual block in region 3, as shown in FIG. 13. [0088] FIG. 14 illustrates Luma down-sampling in region 2 of a neural network in accordance with some implementations of the present disclosure. As shown in FIG. 14, the Y patch, after entering the neural network, is scaled down by a corresponding convolution block or layer in region 2 of the neural network. An inverse operation, e.g., scaling-up, is performed on the corresponding Y patch outputted by the last residual block in region 3, as shown in FIG. 14.

[0089] In the region 3, resolution of the input YUV patches may be scaled up/down at one of the earlier residual blocks, and inverse operations, e.g., scale down/up, may be performed at later residual blocks. As this region is located after YUV concatenation, if scale down operation is performed, input information may be less significantly lost than region 2 since most input information are already captured or learned in earlier convolution layers which have enough depth for information learning. For example, after region 2, three channels of YUV content which has UV scaled up to 128x128 are generated. Y input information may be already leamed/extracted and distributed/duplicated in earlier convolutional layers before the concatenation. Alternatively, a scale down operation may be performed after the first residual block because the first residual block may have enough channels to leam/extract Y input information features. QP independent neural network model

[0090] In order to facilitate easier deployment of the proposed neural network model filtering, it is desired to remove input quantization parameter (QP) dependency from the neural network model. Thus, a single neural network model may be used for image filtering regardless of the input QP used for the video coding.

[0091] FIG. 11 illustrates a typical QP -independent neural network model in accordance with some implementations of the present disclosure. For a typical video coding system, a QP value is used to calculate quantization step size for prediction residual quantization/de-quantization. Therefore, a different QP value represents a different level of video quality. In order to handle different video frames with different input QP and quality, a QpMap is provided to the neural network. The QpMap adds another dimension of information for neural network to learn and adaptively filter the provided YUV input which may include different levels of video quality (e.g., input Qp). In some common video coding standards such as HEVC, VVC or AVS, a predefined relationship (e.g., Q _step = 2^ ^p-4 ^ ⁶ ) is typically used when an input Qp value is converted into a Qp step size for prediction residual quantization. For easier illustration, a QpMap containing or including either input Qp values or Qp step size values is used to introduce the proposed ideas as below.

Dynamic range control of QpMap values

[0092] FIG. 11 illustrates a typical QP-independent neural network based model to perform image filtering for video coding in accordance with some implementations of the present disclosure. The QpMap are concatenated with YUV input channels. Each QpMap channel may contain the same value at different coordinates. And each QpMap channel may have the same resolution as the associated input channel. This is to say, QpMap channel for Y input has the same resolution as the Y input channel and each value of the QpMap indicates that all the samples within the Y input channel has the same Qp values.

[0093] Input QP value of each video frame/image may be directly used to generate QpMap value. Alternatively, input QP value of each video frame/image may be converted into Qp step size first, generate QpMap value.

[0094] When QpMap values are generated from input QP or Qp step size, the QpMap value dynamic range is desired to be reasonable in following three senses. [0095] First, the range should be large enough, so that different QpMap value can be easily used to represent/differentiate different input Qp or Qp step size. In other words, given two input Qp value, the corresponding QpMap values should not be close to each other.

[0096] Second, the range should be balanced enough, so that different QpMap value can evenly distributed at different positions of the range.

[0097] Third, the range should match the dynamic range of the associated YUV sample values. For example, if YUV sample values are normalized to [0, 1] by dividing the P _max , where P _max = 2bitdepth > _t h _e QpM _a p _vaiue i _s supposed to be normalized in a similar range as well.

[0098] Therefore, when QpMap is mapped to a dynamic range, it is proposed to not use the maximum or minimum input Qp or Qp step size as the dividing factor, otherwise the division will push the generated QpMap value towards one side of the dynamic range, which is equal to reduce the dynamic range of the QpMap value.

[0099] For example, if using maximal QpStep size 912 (corresponding to maximal input Qp 63), the theoretic dynamic range is (0, 1), but if input Qp step size is typically less than 45 (corresponding to input Qp 37), the effective dynamic range is only (0, 0.05), which means that in most cases QpMap value is close to 0.

[0100] Instead, it is proposed to use middle/median input Qp or Qp step size to do the normalization, so that the generated QpMap value may be distributed at either side of the dynamic range, e.g. [0.5, 1.5], One exemplary selected middle/median input Qp value is Qp 32, and then converted into Qp step size is approximately 25.5, according to a QP-to-Qstep (Qp step size) equation (e.g., Q _step = 2® ^p-4 ^ ⁶ ). Thus, any input Qp value is first converted to a corresponding Qp step size and then followed by a division by the selected Qp Step size 25.5.

[0101] This selected input Qp or Qp step size for normalization division may be flexibly determined based on the actual input Qp range. For example, given an actual Qp range [22, 42], Qp 37 or its corresponding Qp step size may be selected as the dividing factor such that the normalized value of lower Qps in the actual Qp range are not too close to zero while the normalized value of higher Qps are not too much over 1.0. Alternatively, the maximum value of the actual input QP range (e.g., 42 of the actual Qp range [22, 42]) or its corresponding Qp step size may be selected as the dividing factor if the maximum value is not too big (e.g., within twice size) than the minimum value. Prediction based adjustment of QpMap values

[0102] For inter-predicted video frames/images, most blocks/CTUs in the frame/image may be inter-predicted with small or no residuals, e.g., skip mode. In this case, the effective input Qp value should be determined by the corresponding reference frames/images.

[0103] In some examples, the input Qp value of the corresponding reference frame/image may be saved and obtained when the current image is reconstructed during motion compensation process. The input QP value of each current frame is known. But the input Qp value of this frame becomes unknown when this frame is not current and this frame is a reference frame for another frame. Thus, Qp values have to be saved in order to obtain it in the future.

[0104] In some examples, the input Qp value of the corresponding reference frame/image may be derived by subtracting a specific value from the QP value of the current frame which is inter-coded, where the specific value may be obtained by checking the temporal layer index of the current frame which is inter-coded.

[0105] In some other examples, if the reference frame/image is a chain of reference images (the reference frame/image of a reference frame/image), this information may be inherited or carried over from signaling.

[0106] In a simple solution, for an inter-predicted video frame/image, the effective Qp step size may be derived by a constant scaling factor such as 0.5 from the Qp step size of the current frame, which corresponds to an input Qp difference with value 6. This scaling operation is an approximation to the reference frame/image input Qp or Qp step size.

[0107] In one or more examples, the scaling of Qp Step size may be applied to an intra-predicted frame/image as well, to compensate the inaccuracy of the constant scaling factor used for subsequent inter-predicted frames.

[0108] In another example, the Qp scaling factor may be flexibly derived using following different methods.

[0109] In the first method, the Qp scaling factor may be selected by the encoder from a set of values. The scaling factor set may be sequence based or image/slice based, which means the set may be coded in the picture header or sequence parameter set. The index of the selected Qp scaling factor in the scaling factor set may be selected based on rate-distortion optimization algorithms at the encoder side with different granularities, e.g., picture level index selection, CTU level index selection, block level selection (e.g., a picture may be portioned into different blocks based on quad-tree division), for the purpose of good trade-off between image quality and signaling overhead.

[0110] In the second method, the Qp scaling factor may be converted to a Qp offset/adjustment. The Qp offset/adjustment may be applied to the input Qp or Qp step size before calculating the QpMap values.

[OHl] In the third method, for current CTU/picture/block, the Qp scaling factor may be inherited from adjacent CTUs/picture/blocks (e.g., left or above CTUs/blocks in spatial domain, or the reference blocks/CTUs in time domain) to save signaling overhead.

[0112] In the fourth method, instead of signaling or inheritance, the Qp scaling factor may be calculated at the decoder side by the Qp difference of the current CTU/block and reference CTU/block. If there are multiple reference CTUs/blocks corresponding to the current CTU/block, an averaged value of the Qp scaling factor may be calculated. If a reference CTU/block is involved in a reference chain. The reference depth may be constrained and the Qp scaling factor may be calculated according to the parent reference CTU/block at most of the constrained reference depth. [0113] In the fifth method, the Qp scaling factor of different components may be jointly signaled/selected/calculated for lower complexity. Alternatively, the Qp scaling factor of different components may be separately signaled/selected/calculated. Alternatively, the Qp scaling factor of luma and chroma may be separately signaled/selected/calculated.

[0114] In the sixth method, any combination of the above methods may be used as a hybrid method. QpMap value based sample value scaling

[0115] A QP-independent neural network model may not explicitly contain QpMap channels in the network. For example, as shown in FIG. 11, the QpMap value generated for each YUV channel is concatenated with YUV channels. Alternatively, the QpMap value fed into the network may be used to directly scale the sample values in each YUV channel. In this way, QpMap channels are not concatenated with YUV channels, which represents an implicit use of QpMap in the network. [0116] In some examples of the implicit use of QpMap, the QpMap data may not be fed into the network, the scaling of sample values in each YUV channel is performed before the neural network. Interactions between neural network based model filtering and other in-loop filters

[0117] When QpMap channels are provided to a neural network to filter video content with different qualities, the QpMap channels may include Qp information from one or more components. [0118] FIG. 12A illustrate an example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure. FIG. 12B illustrate another example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure. In FIGS. 12A-12B, block Map-Y indicates QpMap channel for Y channel, block Map-U indicates QpMap channel for U channel, and block Map-V indicates QpMap channel for V channel. Blocks Y, U, and V respectively indicate Y channel, U channel, and V channel.

[0119] Given a YUV420 content, the UV components may be up-sampled first and then YUV are collocated and interleaved with corresponding QpMap channels as shown in FIG. 12A, or the Y channel may be down-sampled into four smaller Y channels first and then YUV are collocated and interleaved with QpMap channels as shown in FIG. 12B. In some examples, either the up-sampling or down-sampling is performed within the network, e.g., regions 2 and 3 in FIG. 10 or outside of the network, e.g., region 1 in FIG. 10.

[0120] FIG. 16A illustrates another example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure. FIG. 16B illustrates another example of layout of collocated QpMap channels and YUV channels in accordance with some implementations of the present disclosure. In FIGS. 16A-16B, block Map-Y indicates QpMap channel for Y channel, block Map-U indicates QpMap channel for U channel, and block Map-V indicates QpMap channel for V channel. Blocks Y, U, and V respectively indicate Y channel, U channel, and V channel. The multiple QpMap channels may be concatenated internally first then concatenated with YUV channels, as shown in FIGS. 16A-16B.

[0121] In case only one or more QpMap channels of one component is/are provided to a neural network, the one or more QpMap channels may be positioned on one side of the YUV channels, which indicates that the YUV channels are concatenated before the addition of QpMap channels so that YUV channels are adjacently collocated.

[0122] In another example, in addition of the QpMpa channels from different components, additional QpMap channels for different types of training data may be needed. For example, if the training data is cropped from I frames or B frames or P frames, a QpMap containing frame type information may be generated and concatenated. The I frames are intra-coded frames, the B frames are bi-directional predicted frames, and the P frames are predicted frames. Filtering offset or scaling of the output of the neural network model based filter

[0123] For generalization purpose, a unified neural network model based filter may be used for different video content with different level of qualities, motion and illumination environment. The output of the neural network model based filter may be slightly adjusted in the form of offset or scaling at the encoder side for better coding efficiency.

[0124] The filtering offset or scaling value may be adaptively selected by the encoder from a set of values. The offset or scaling set may be sequence based or image/slice based, which means the set may be coded in the picture/slice header or sequence parameter set. The index of the selected offset or scaling value in the set may be selected based on rate-distortion optimization algorithms at the encoder side with different granularities, e.g., picture level index selection, CTU level index selection, block level selection, e.g., a picture may be portioned into different blocks based on quad-tree division, for the purpose of good trade-off between image quality and signaling overhead. [0125] The selection of the adaptive filtering offset or scaling value may be based on certain classification algorithms such as content smoothness or histogram of oriented gradients. The adaptive filtering offset or scaling value of each category is calculated and selected at the encoder and explicitly signaled to the decoder for reducing sample distortion effectively, while the classification of each sample is performed at both the encoder and the decoder for saving side information significantly.

[0126] The selection of the adaptive filtering offset or scaling value may be jointly or separately performed for different components, e.g., YUV have different adaptive filtering offset or scaling values.

Training data generation and training process

[0127] When a neural network based filter model is trained, the training data preparation and training process may be performed in different ways.

[0128] In some examples, the model may be trained based on a data set with only still images. The data set may be encoded with all I frames from the video coding tool where the neural network based filter is used.

[0129] In some examples, the model may be trained based on a two-path process. In the first path, a data set may be encoded with all I frames and a model A may be trained based on all the I frames. In the second path, the same data set or a new data set may be encoded with a combination of I, B and P frames with different ratios (the number ratio of the contained I, B and P frames). In some examples, the generated I/B/P frames are encoded by applying the model A trained in the first path. Based on the newly generated I/B/P frames, a new model B may be trained.

[0130] When model B is trained, model A may be loaded as a pre-trained model such that model B is a refined model staring from model A. In another example, another model different from model A may be loaded as a pre-trained point.

[0131] Alternatively, model B may be trained from scratch.

[0132] In some examples, the model may be trained multi-path that is more than two-path. In the first path, a model A may be trained based on I frames. In the second path, a model B may be trained or refined based on model A based on a combination of I/B/P frames when model A is applied to the encoder. Note that the selected combination of the B/P frames during this second training path may be only from low temporary layers. In the third path or further paths, a model C may be trained or refined based on model B based on higher temporary layers of B/P frames. When higher temporary layers of B/P frames are generated and selected, model B or/and model A may be applied at the encoder side.

[0133] Before network training, training data must be generated. In this multi-path method, training data is generated by three paths including: the first pass is to generate I frames only, which is used to train model A; once model A is ready, the encoder may or may not load model A and generate low temporal layer B/P frames, which is called the second path. These generated low temporal layer B/P frames are used to train model B by a new training or being refined based on model A.

[0134] Further, once model B is ready, the encoder may or may not load model A and B and generate high temporal layer B/P frames, which is called the third path. These generated high temporal layer B/P frames are used to train model C by a new training or being refined based on model A or/and B.

Interactions between neural network based model filtering and other in-loop filters

[0135] When neural network based model filtering is signaled to be turned on at CTU level or frame level, the deblocking filtering may be skipped to avoid unnecessary computation or oversmoothing. Alternatively, the deblocking filtering may be still performed for visual quality purpose. [0136] When neural network based model filtering is signaled to be turned on at CTU level or frame level, some other in-loop filters such ALF, Cross Component Adaptive Loop Filter (CCALF) and SAO may be turned off. [0137] When neural network based model filtering is signaled to be turned on at CTU level or frame level, other in-loop filters may be selectively turned on or off at CTU level or frame level. For example, if an intra-frame or an intra-frame CTU is enabled for neural network based model filtering, other in-loop filters such as deblocking filtering, or/and ALF, or/and CCALF, or/and SAO, for current intra-frame, or current intra-frame CTU are disabled.

[0138] FIG. 17 is a block diagram illustrating an apparatus for image filtering in video coding using a neural network in accordance with some implementations of the present disclosure. The apparatus 1700 may be a terminal, such as a mobile phone, a tablet computer, a digital broadcast terminal, a tablet device, or a personal digital assistant.

[0139] As shown in FIG. 17, the apparatus 1700 may include one or more of the following components: a processing component 1702, a memory 1704, a power supply component 1706, a multimedia component 1708, an audio component 1710, an input/output (VO) interface 1712, a sensor component 1714, and a communication component 1716.

[0140] The processing component 1702 usually controls overall operations of the apparatus 1700, such as operations relating to display, a telephone call, data communication, a camera operation, and a recording operation. The processing component 1702 may include one or more processors 1720 for executing instructions to complete all or a part of steps of the above method. Further, the processing component 1702 may include one or more modules to facilitate interaction between the processing component 1702 and other components. For example, the processing component 1702 may include a multimedia module to facilitate the interaction between the multimedia component 1708 and the processing component 1702.

[0141] The memory 1704 is configured to store different types of data to support operations of the apparatus 1700. Examples of such data include instructions, contact data, phonebook data, messages, pictures, videos, and so on for any application or method that operates on the apparatus 1700. The memory 1704 may be implemented by any type of volatile or non-volatile storage devices or a combination thereof, and the memory 1704 may be a Static Random Access Memory (SRAM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), an Erasable Programmable Read-Only Memory (EPROM), a Programmable Read-Only Memory (PROM), a Read-Only Memory (ROM), a magnetic memory, a flash memory, a magnetic disk or a compact disk. [0142] The power supply component 1706 supplies power for different components of the apparatus 1700. The power supply component 1706 may include a power supply management system, one or more power supplies, and other components associated with generating, managing and distributing power for the apparatus 1700.

[0143] The multimedia component 1708 includes a screen providing an output interface between the apparatus 1700 and a user. In some examples, the screen may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen receiving an input signal from a user. The touch panel may include one or more touch sensors for sensing a touch, a slide and a gesture on the touch panel. The touch sensor may not only sense a boundary of a touching or sliding actions, but also detect duration and pressure related to the touching or sliding operation. In some examples, the multimedia component 1708 may include a front camera and/or a rear camera. When the apparatus 1700 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera may receive external multimedia data.

[0144] The audio component 1710 is configured to output and/or input an audio signal. For example, the audio component 1710 includes a microphone (MIC). When the apparatus 1700 is in an operating mode, such as a call mode, a recording mode and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal may be further stored in the memory 1704 or sent via the communication component 1716. In some examples, the audio component 1710 further includes a speaker for outputting an audio signal.

[0145] The I/O interface 1712 provides an interface between the processing component 1702 and a peripheral interface module. The above peripheral interface module may be a keyboard, a click wheel, a button, or the like. These buttons may include but not limited to, a home button, a volume button, a start button, and a lock button.

[0146] The sensor component 1714 includes one or more sensors for providing a state assessment in different aspects for the apparatus 1700. For example, the sensor component 1714 may detect an on/off state of the apparatus 1700 and relative locations of components. For example, the components are a display and a keypad of the apparatus 1700. The sensor component 1714 may also detect a position change of the apparatus 1700 or a component of the apparatus 1700, presence or absence of a contact of a user on the apparatus 1700, an orientation or acceleration/decel eration of the apparatus 1700, and a temperature change of apparatus 1700. The sensor component 1714 may include a proximity sensor configured to detect presence of a nearby object without any physical touch. The sensor component 1714 may further include an optical sensor, such as a CMOS or CCD image sensor used in an imaging application. In some examples, the sensor component 1714 may further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0147] The communication component 1716 is configured to facilitate wired or wireless communication between the apparatus 1700 and other devices. The apparatus 1700 may access a wireless network based on a communication standard, such as WiFi, 4G, or a combination thereof. In an example, the communication component 1716 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an example, the communication component 1716 may further include a Near Field Communication (NFC) module for promoting short-range communication. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, Ultra-Wide Band (UWB) technology, Bluetooth (BT) technology and other technology.

[0148] In an example, the apparatus 1700 may be implemented by one or more of Application Specific Integrated Circuits (ASIC), Digital Signal Processors (DSP), Digital Signal Processing Devices (DSPD), Programmable Logic Devices (PLD), Field Programmable Gate Arrays (FPGA), controllers, microcontrollers, microprocessors, or other electronic elements to perform the above method. A non-transitory computer readable storage medium may be, for example, a Hard Disk Drive (HDD), a Solid-State Drive (SSD), Flash memory, a Hybrid Drive or Solid-State Hybrid Drive (SSHD), a Read-Only Memory (ROM), a Compact Disc Read-Only Memory (CD-ROM), a magnetic tape, a floppy disk, etc.

[0149] FIG. 18 is a flowchart illustrating a process for image filtering in video coding using a neural network in accordance with some implementations of the present disclosure.

[0150] In step 1801, the processor 1720 loads a plurality of QpMap values at a plurality of QpMap channels into the neural network. As shown in FIG. 11, the plurality of QpMap channels respectively combined with corresponding YUV channels after the three convolution blocks or layers before the concatenation block.

[0151] For example, QpMap may have three channels for YUV respectively: QP-Y, QP-U, QP-V because YUV may have different QP values. [0152] In step 1802, the processor 1720 adjusts, according to a QP scaling factor, the plurality of QpMap values for the neural network to learn and filter a current input frame to the neural network [0153] In some examples, an encoder may select the QP scaling factor from a plurality of values, where the plurality of values are coded in a picture header or a sequence parameter set associate the current input frame.

[0154] In some examples, the processor 1720 may convert the QP scaling factor to a QP offset and apply the QP offset to input QP values of the current input frame before calculating the plurality of QpMap values.

[0155] In some examples, the processor 1720 may inherit the QP scaling factor from inheriting the QP scaling factor from an adjacent element to a current input to the neural network, where the adjacent element may include a coding tree unit, a picture, a frame, or a block. For example, the QP scaling factor may be inherited from different granularities. The Qp scaling factor may be inherited from adjacent CTUs/picture/blocks, e.g., left or above CTUs/blocks in spatial domain, or the reference blocks/CTUs in time domain, to save signaling overhead.

[0156] In some examples, the QP scaling factor may be calculated based on a QP difference between a current input to the neural network and a reference element at different granularities. The reference element may include reference CTU/block. For example, the Qp scaling factor may be calculated at the decoder side by the Qp difference of the current CTU/block and reference CTU/block. If there are multiple reference CTUs/blocks corresponding to the current CTU/block, an averaged value of the Qp scaling factor may be calculated. If a reference CTU/block is involved in a reference chain. The reference depth may be constrained and the Qp scaling factor may be calculated according to the parent reference CTU/block at most of the constrained reference depth. [0157] In some examples, the processor 1720 may jointly select the QP scaling factor for a plurality of input patches associated with the current input frame that are loaded to the neural network at different input channels.

[0158] In some examples, the processor 1720 may load a plurality of input patches associated with the current input frame into the neural network at a plurality of input channels and combine the plurality of QpMap channels with the plurality of input channels. FIGS. 12A-12B and 16A-16B respectively illustrate different layout for combining the plurality of input channels and the plurality of QpMap channels. [0159] In some examples, the plurality of input patches may include a plurality of first input patches loaded at a first input channel, a plurality of second input patches loaded at a second input channel, a plurality of third input patches loaded at a third input channel; the plurality of QpMap channels may include a first QpMap channel for the first input channel, a second QpMap channel for the second input channel, and a third QpMap channel for the third input channel. The processor 1720 may obtain an adjusted second input channel and an adjusted third input channel by respectively up-sampling the second input channel and the third input channel and interleave the first QpMap channel, the second QpMap channel, and the third QpMap channel with the first input channel, the adjusted second input channel, and the adjusted third channel, as shown in FIG. 12A. [0160] In some examples, the plurality of input patches may include a plurality of first input patches loaded at a first input channel, a plurality of second input patches loaded at a second input channel, and a plurality of third input patches loaded at a third input channel. The processor 1720 may obtain a plurality of adjusted first input channels by down-sampling the first input channel, where the plurality of QpMap channels include a plurality of first QpMap channels for the plurality of adjusted first input channels, a second QpMap channel for the second input channel, and a third QpMap channel for the third input channel; and interleave the plurality of first QpMap channels, the second QpMap channel, and the third QpMap channel with the plurality of adjusted first input channels, the second input channel, and the third input channel.

[0161] In some examples, the plurality of input patches may include a plurality of first input patches loaded at a first input channel, a plurality of second input patches loaded at a second input channel, a plurality of third input patches loaded at a third input channel. The processor 1720 may obtain a concatenated QpMap channel by concatenating the plurality of QpMap channels and concatenate the concatenated QpMap channel, the first input channel, the second input channel, and the third input channel.

[0162] In some examples, the the processor 1720 may obtain an adjusted second input channel and an adjusted third input channel by respectively up-sampling the second input channel and the third input channel and concatenate the concatenated QpMap channel, the first input channel, the second input channel, and the third input channel, where the plurality of QpMap channels include a first QpMap channel for the first input channel, a second QpMap channel for the adjusted second input channel, and a third QpMap channel for the adjusted third input channel, as shown in FIG. 16A. In some examples, the the processor 1720 may obtain a plurality of adjusted first input channels by down-sampling the first input channel, where the plurality of QpMap channels include a plurality of first QpMap channels for the plurality of adjusted first input channels, a second QpMap channel for the second input channel, and a third QpMap channel for the third input channel; and concatenate the concatenated QpMap channel, the plurality of adjusted first input channels, the second input channel, and the third input channel, as shown in FIG. 16B.

[0163] FIG. 19 is a flowchart illustrating a process for adjusting outputs of a neural network based image filter in video coding in accordance with some implementations of the present disclosure.

[0164] In step 1901, the processor 1720 adaptively selects an adjustment factor from a plurality of values, where the plurality of values are coded in a picture header or a sequence parameter set associate a current input frame to be coded. The processor 1720 may be applied in an encoder.

[0165] In step 1902, the processor 1720 adjusts, according to the adjustment factor, an output of the neural network based image filter.

[0166] In some examples, the adjustment factor may include a filtering offset or a scaling factor.

[0167] In some examples, the processor 1720 may select the adjustment factor based on ratedistortion optimization with different granularities.

[0168] In some examples, the processor 1720 may determine a classification of the current input frame and select the adjustment factor based on a classification algorithm and signaling the adjustment factor such that a decoder determines the classification and obtains the adjustment factor.

[0169] In some examples, the output of the neural network based image filter may include a plurality of patches at different channels. The processor 1720 may jointly or separately select the adjustment factor for each of the plurality of patches at the different channels.

[0170] FIG. 20 is a flowchart illustrating an example process for training a neural network based image filter in accordance with the present disclosure.

[0171] In step 2001, the processor 1720 trains the neural network based image filter according to a plurality of intra-coded frames to obtain a first model and a first data set encoded with the plurality of intra-coded frames. In some examples, the first data set may consist of still images.

[0172] In step 2002, the processor 1720 generates a plurality of I/B/P frames including one or more intra-coded frames, one or more bi-directional predicted frames, and one or more predicted frames. [0173] In step 2003, the processor 1720 obtains a second data set by encoding the plurality of I/B/P pictures using the first model. The first model may be model A.

[0174] In step 2004, the processor 1720 obtains a second model based on the second data set. The second model may be model B.

[0175] In some examples, the processor 1720 may obtain the second model by loading the first model as a pre-trained model.

[0176] In some examples, the processor 1720 may select the plurality of I/B/P frames from low temporary layers.

[0177] In some examples, the processor 1720 may obtain a third model by training the second model based on high temporary layers of the bi-directional predicted frames or the predicted frames. The third model may be model C.

[0178] In some other examples, there is provided a non-transitory computer readable storage medium 1704, having instructions stored therein. When the instructions are executed by one or more processors 1720, the instructions cause the processor to perform any method as described in FIGS. 18-20 and above.

[0179] The description of the present disclosure has been presented for purposes of illustration and is not intended to be exhaustive or limited to the present disclosure. Many modifications, variations, and alternative implementations will be apparent to those of ordinary skill in the art having the benefit of the teachings presented in the foregoing descriptions and the associated drawings.

[0180] The examples were chosen and described in order to explain the principles of the disclosure, and to enable others skilled in the art to understand the disclosure for various implementations and to best utilize the underlying principles and various implementations with various modifications as are suited to the particular use contemplated. Therefore, it is to be understood that the scope of the disclosure is not to be limited to the specific examples of the implementations disclosed and that modifications and other implementations are intended to be included within the scope of the present disclosure.