Neural network based video coding tools

Abstract

A coding method is implemented by a decoding device in a convolutional neural network (CNN). The coding method includes receiving transform coefficients. The method also includes applying a quantization step (QS) to the transform coefficients during a quantization process instead of applying a base quantization parameter (QP) or a slice QP to obtain a reconstructed transform. The method also includes generating an image based on the reconstructed transform.

Description

Neural Network Based Video Coding ToolsCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This patent application claims the benefit of U.S. Provisional Patent Application No. 63 / 634,035 filed April 15, 2024, and U.S. Provisional Patent Application No. 63 / 634,041 filed April 15, 2024, both of which are hereby incorporated by reference.TECHNICAL FIELD

[0002] In general, this disclosure describes techniques for image reconstruction. More specifically, this disclosure ensures improvement to reconstruction accuracy of inverse transforms using neural network-based video coding tools.BACKGROUND

[0003] The amount of video data needed to depict even a relatively short video can be substantial, which may result in difficulties when the data is to be streamed or otherwise communicated across a communications network with limited bandwidth capacity. Thus, video data is generally compressed before being communicated across modern day telecommunications networks. The size of a video could also be an issue when the video is stored on a storage device because memory resources may be limited. Video compression devices often use software and / or hardware at the source to code the video data prior to transmission or storage, thereby decreasing the quantity of data needed to represent digital video images. The compressed data is then received at the destination by a video decompression device that decodes the video data. With limited network resources and ever-increasing demands of higher video quality, improved compression and decompression techniques that improve compression ratio with little to no sacrifice in image quality are desirable.SUMMARY

[0004] A first aspect relates to a method of decoding a coded video bitstream implemented by a decoding device in a convolutional neural network (CNN). The method includes receiving, by the decoding device, transform coefficients; applying, by the decoding device, a quantization step (QS) to the transform coefficients during a quantization process instead of applying a basequantization parameter (QP) or a slice QP to obtain a reconstructed transform; and generating, by the decoding device, an image based on the reconstructed transform.

[0005] The method provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level. In particular, a QS is applied as input during a quantization process instead of applying QPs. By applying QS to control the quantization level, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0006] Optionally, in any of the preceding aspects, another implementation of the aspect provides that applying the QS includes applying, by the decoding device, an activation function during the quantization process.

[0007] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the activation function includes a concatenated rectified linear unit (CReLU).

[0008] Optionally, in any of the preceding aspects, another implementation of the aspect provides applying, by the decoding device, the CReLU instead of a rectified linear unit (ReLU) or a parametric ReLU (PReLU).

[0009] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0010] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0011] A second aspect relates to a method of encoding a coded video bitstream implemented by a decoding device in a convolutional neural network CNN. The method includes receiving, bythe encoding device, transform coefficients; applying, by the encoding device, a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform; and generating, by the encoding device, bitstream based on the reconstructed transform.

[0012] The method provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level. In particular, a QS is applied as input during a quantization process instead of applying QPs. By applying QS to control the quantization level, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0013] Optionally, in any of the preceding aspects, another implementation of the aspect provides that applying the QS includes applying, by the encoding device, an activation function during the quantization process.

[0014] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the activation function includes a CReLU.

[0015] Optionally, in any of the preceding aspects, another implementation of the aspect provides applying, by the encoding device, the CReLU instead of a ReLU or a PReLU.

[0016] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0017] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0018] A third aspect relates to a computing device. The computing device includes means for receiving transform coefficients; means for receiving transform coefficients; means for applying a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform; and means for generating an image based on the reconstructed transform.

[0019] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level. In particular, a QS is applied as input during a quantization process instead of applying QPs. By applying QS to control the quantization level, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0020] Optionally, in any of the preceding aspects, another implementation of the aspect provides that means for applying the QS includes means for applying an activation function during the quantization process.

[0021] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the activation function includes a CReLU.

[0022] Optionally, in any of the preceding aspects, another implementation of the aspect provides means for applying the CReLU instead of a ReLU or a PReLU.

[0023] A fourth aspect relates to a computing device. The computing device includes means for receiving transform coefficients; means for receiving transform coefficients; means for applying a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform; and means for generating a bitstream based on the reconstructed transform.

[0024] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level. In particular, a QS is applied as input during a quantization process instead of applying QPs. By applying QS to control the quantization level, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / ornetwork resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0025] Optionally, in any of the preceding aspects, another implementation of the aspect provides that means for applying the QS includes means for applying an activation function during the quantization process.

[0026] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the activation function includes a CReLU.

[0027] Optionally, in any of the preceding aspects, another implementation of the aspect provides means for applying the CReLU instead of a ReLU or a PReLU.

[0028] A fifth aspect relates to a coding method implemented by a decoding device in a CNN. The method includes: applying an enhancement to transform coefficients; generating enhanced transform coefficients based on the enhancement; applying an inverse transform to the enhanced transform coefficients to generate a reconstructed transform; and generating an image based on the reconstructed transform.

[0029] The method provides techniques that improve reconstruction accuracy of inverse transforms by applying an enhancement process to inverse transform coefficients. In particular, a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0030] Optionally, in any of the preceding aspects, another implementation of the aspect provides that applying the enhancement includes applying a plurality of enhancement processes, and where applying the inverse transform includes applying a plurality of inverse transform operations.

[0031] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the method includes performing an enhancement process before the inverse transform is applied.

[0032] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the plurality of enhancement processes includes a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and that the plurality of inverse transform operations include an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

[0033] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the method includes performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

[0034] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0035] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0036] A sixth aspect relates to a coding method implemented by an encoding device in a CNN. The method includes: applying an enhancement to transform coefficients; generating enhanced transform coefficients based on the enhancement; applying an inverse transform to the enhanced transform coefficients to generate a reconstructed transform; and generating a bitstream based on the reconstructed transform.

[0037] The method provides techniques that improve reconstruction accuracy of inverse transforms by applying an enhancement process to inverse transform coefficients. In particular, acoefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0038] Optionally, in any of the preceding aspects, another implementation of the aspect provides that applying the enhancement includes applying a plurality of enhancement processes, and where applying the inverse transform includes applying a plurality of inverse transform operations.

[0039] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the method includes performing an enhancement process before the inverse transform is applied.

[0040] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the plurality of enhancement processes includes a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and that the plurality of inverse transform operations include an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

[0041] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the method includes performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

[0042] A seventh aspect relates to a computing device. The computing device includes means for applying an enhancement to transform coefficients; means for generating enhanced transform coefficients based on the enhancement; means for applying an inverse transform to the enhanced transform coefficients to generate a reconstructed transform; and means for generating an image based on the reconstructed transform.

[0043] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by applying an enhancement process to inverse transform coefficients. In particular, a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0044] Optionally, in any of the preceding aspects, another implementation of the aspect provides that means for applying the enhancement includes means for applying a plurality of enhancement processes, and where means for applying the inverse transform includes applying a plurality of inverse transform operations.

[0045] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the computing device includes means for performing an enhancement process before the inverse transform is applied.

[0046] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the plurality of enhancement processes includes a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and that the plurality of inverse transform operations include an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

[0047] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the computing device includes performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

[0048] An eighth aspect relates to a computing device. The computing device includes means for applying an enhancement to transform coefficients; means for generating enhanced transform coefficients based on the enhancement; means for applying an inverse transform to the enhancedtransform coefficients to generate a reconstructed transform; and means for generating a bitstream based on the reconstructed transform.

[0049] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by applying an enhancement process to inverse transform coefficients. In particular, a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0050] Optionally, in any of the preceding aspects, another implementation of the aspect provides that means for applying the enhancement includes means for applying a plurality of enhancement processes, and where means for applying the inverse transform includes applying a plurality of inverse transform operations.

[0051] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the computing device includes means for performing an enhancement process before the inverse transform is applied.

[0052] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the plurality of enhancement processes includes a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and that the plurality of inverse transform operations include an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

[0053] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the computing device includes performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

[0054] A ninth aspect relates to a method of coding implemented by a decoding device in a CNN. The method includes: inputting a reconstructed block into a convolution layer; applying alow-rank decomposition to the reconstructed block within the convolution layer; generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and generating an image based on the filtered reconstructed block.

[0055] The method provides techniques that improve reconstruction accuracy of inverse transforms by applying a low-rank decomposition to a convolutional layer. In particular, the method decomposes a convolutional layer to produce a convolutional layer having a fewer number of filters. By applying the low-rank decomposition, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0056] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the reconstructed block includes a first matrix that is output from a previous layer, that the method further includes applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and that the second matrix has a lower weight than the first matrix.

[0057] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the first matrix includes a first number of coefficients and the second matrix includes a second number of coefficients, and that the second number is less than the first number.

[0058] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0059] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0060] A tenth aspect relates to a method of coding implemented by an encoding device in a CNN. The method includes: inputting a reconstructed block into a convolution layer; applying a low-rank decomposition to the reconstructed block within the convolution layer; generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and generating a bitstream based on the filtered reconstructed block.

[0061] The method provides techniques that improve reconstruction accuracy of inverse transforms by applying a low-rank decomposition to a convolutional layer. In particular, the method decomposes a convolutional layer to produce a convolutional layer having a fewer number of filters. By applying the low-rank decomposition, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0062] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the reconstructed block includes a first matrix that is output from a previous layer, that the method further includes applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and that the second matrix has a lower weight than the first matrix.

[0063] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the first matrix includes a first number of coefficients and the second matrix includes a second number of coefficients, and that the second number is less than the first number.

[0064] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0065] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by oneor more processors, cause the computing device to execute the method in any of the preceding aspects.

[0066] An eleventh aspect relates to a computing device. The computing device includes: means for inputting a reconstructed block into a convolution layer; means for applying a low-rank decomposition to the reconstructed block within the convolution layer; means for generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and means for generating an image based on the filtered reconstructed block.

[0067] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by applying a low-rank decomposition to a convolutional layer. In particular, the computing device decomposes a convolutional layer to produce a convolutional layer having a fewer number of filters. By applying the low-rank decomposition, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0068] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the reconstructed block includes a first matrix that is output from a previous layer, that the computing device further includes means for applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and that the second matrix has a lower weight than the first matrix.

[0069] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the first matrix includes a first number of coefficients and the second matrix includes a second number of coefficients, and that the second number is less than the first number.

[0070] A twelfth aspect relates to a computing device. The computing device includes: means for inputting a reconstructed block into a convolution layer; means for applying a low-rank decomposition to the reconstructed block within the convolution layer; means for generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and means for generating a bitstream based on the filtered reconstructed block.

[0071] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by applying a low-rank decomposition to a convolutional layer. In particular, the computing device decomposes a convolutional layer to produce a convolutional layer having a fewer number of fdters. By applying the low-rank decomposition, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0072] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the reconstructed block includes a first matrix that is output from a previous layer, that the computing device further includes means for applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and that the second matrix has a lower weight than the first matrix.

[0073] Optionally, in any of the preceding aspects, another implementation of the aspect provides that the first matrix includes a first number of coefficients and the second matrix includes a second number of coefficients, and that the second number is less than the first number.

[0074] A thirteenth aspect relates to a method of decoding a coded video bitstream implemented by a decoding device in a CNN. The method includes receiving, by the decoding device, transform coefficients; generating, by the decoding device, a reconstructed transform by: applying a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP; applying an enhancement to the transform coefficients; and applying an inverse transform to the enhanced transform coefficients. The decoding device generates an image based on the reconstructed transform.

[0075] The method provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level and by applying an enhancement process to inverse transform coefficients. In particular, a QS is applied as input during a quantization process instead of applying QPs, and a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying QS to control the quantization level and by applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inversetransforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0076] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0077] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0078] A fourteenth aspect relates to a method of encoding a coded video bit stream implemented by a encoding device in a CNN. The method includes receiving, by the encoding device, transform coefficients; generating, by the encoding device, a reconstructed transform by: applying a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP; applying an enhancement to the transform coefficients; and applying an inverse transform to the enhanced transform coefficients. The encoding device generates a bitsteam based on the reconstructed transform.

[0079] The method provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level and by applying an enhancement process to inverse transform coefficients. In particular, a QS is applied as input during a quantization process instead of applying QPs, and a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying QS to control the quantization level and by applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As apractical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0080] Optionally, in any of the preceding aspects, another implementation of the aspect provides a coding device including: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of the preceding aspects.

[0081] Optionally, in any of the preceding aspects, another implementation of the aspect provides a non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of the preceding aspects.

[0082] A fifteenth aspect relates to a computing device. The computing device includes means for receiving transform coefficients; means for generating a reconstructed transform including: means for applying a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP; means for applying an enhancement to the transform coefficients; and means for applying an inverse transform to the enhanced transform coefficients. The computing device further includes means for generating an image based on the reconstructed transform.

[0083] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level and by applying an enhancement process to inverse transform coefficients. In particular, a QS is applied as input during a quantization process instead of applying QPs, and a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying QS to control the quantization level and by applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0084] A sixteenth aspect relates to a computing device. The computing device includes means for receiving transform coefficients; means for generating a reconstructed transform including: means for applying a QS to the transform coefficients during a quantization process instead ofapplying a base QP or a slice QP; means for applying an enhancement to the transform coefficients; and means for applying an inverse transform to the enhanced transform coefficients. The computing device further includes means for generating a bistream based on the reconstructed transform.

[0085] The computing device provides techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level and by applying an enhancement process to inverse transform coefficients. In particular, a QS is applied as input during a quantization process instead of applying QPs, and a coefficient enhancement is applied to transform coefficients before applying an inverse transform. By applying QS to control the quantization level and by applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0086] For the purpose of clarity, any one of the foregoing embodiments may be combined with any one or more of the other foregoing embodiments to create a new embodiment within the scope of the present disclosure.

[0087] These and other features will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings and claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0088] For a more complete understanding of this disclosure, reference is now made to the following brief description, taken in connection with the accompanying drawings and detailed description, wherein like reference numerals represent like parts.

[0089] FIG. 1 is a flowchart of an example method of coding a video signal.

[0090] FIG. 2 is a schematic diagram of an example coding and decoding (codec) system for video coding.

[0091] FIG. 3 is a schematic diagram illustrating an example video encoder.

[0092] FIG. 4 is a schematic diagram illustrating an example video decoder.

[0093] FIG. 5 illustrates an embodiment of a coding method implemented by an encoding device.

[0094] FIG. 6 illustrates an embodiment of a coding method implemented by a decoding device.

[0095] FIG. 7 illustrates an embodiment of an inverse transform enhancement.

[0096] FIG. 8 illustrates an embodiment of an inverse primary transform enhancement implemented by a coding device.

[0097] FIG. 9 illustrates an embodiment of an inverse secondary transform enhancement implemented by a coding device.

[0098] FIG. 10 is an example of a neural network-based loop filter.

[0099] FIG. 11A and FIG. 11B illustrate embodiments of layers applied in a neural networkbased loop filter.

[0100] FIG. 12 illustrates an embodiment of a quantization step applied in a coding method implemented by an encoding device.

[0101] FIG. 13 illustrates an embodiment of a quantization step applied in a coding method implemented by a decoding device.

[0102] FIG. 14 illustrates an embodiment of a quantization step and inverse transform enhancement applied in a coding method implemented by an encoding device.

[0103] FIG. 15 illustrates an embodiment of a quantization step and inverse transform enhancement applied in a coding method implemented by a decoding device.

[0104] FIG. 16 illustrates an embodiment of layers applied in a neural network-based loop filter.

[0105] FIG. 17 illustrates an embodiment of a coding method implemented by an encoding device.

[0106] FIG. 18 illustrates an embodiment of a coding method implemented by an decoding device.

[0107] FIG. 19 illustrates an embodiment of a low-rank decomposition applied in a coding method implemented by an encoding device.

[0108] FIG. 20 illustrates an embodiment of a low-rank decomposition applied in a coding method implemented by a decoding device.

[0109] FIG. 21 is a schematic diagram of a video coding device.DETAILED DESCRIPTION

[0110] It should be understood at the outset that although an illustrative implementation of one or more embodiments are provided below, the disclosed systems and / or methods may be implemented using any number of techniques, whether currently known or in existence. The disclosure should in no way be limited to the illustrative implementations, drawings, and techniquesillustrated below, including the exemplary designs and implementations illustrated and described herein, but may be modified within the scope of the appended claims along with their full scope of equivalents.

[0111] The following terms are defined as follows unless used in a contrary context herein. Specifically, the following definitions are intended to provide additional clarity to the present disclosure. However, terms may be described differently in different contexts. Accordingly, the following definitions should be considered as a supplement and should not be considered to limit any other definitions of descriptions provided for such terms herein.

[0112] A bitstream is a sequence of bits including video data compressed for transmission between an encoder and a decoder. An encoder is a device configured to employ encoding processes to compress video data into a bitstream. A decoder is a device configured to employ decoding processes to reconstruct video data (e g., an image) from a bitstream for display. A picture is an array of luma samples and / or an array of chroma samples that create a frame or a field thereof. A picture that is being encoded or decoded can be referred to as a current picture for clarity of discussion. Also for clarity of discussion, an encoder or a decoder may also be referred to as a coding device.

[0113] A neural network is a method in artificial intelligence (Al) that teaches computers to process data in a way that is inspired by the human brain. It is a type of machine learning (ML) process, called deep learning, that uses interconnected nodes or neurons in a layered structure that resembles the human brain. Feedforward neural networks process data in one direction, from the input node to the output node. Every node in one layer is connected to every node in the next layer. A feedforward network uses a feedback process to improve predictions over time. A convolutional neural network (CNN) is a type of feedforward neural network that learns features via filter (or kernel) optimization. This type of deep learning network has been applied to process and make predictions from many different types of data including text, images and audio.

[0114] A convolutional layer is a building block of CNNs that applies a convolution operation to the input, using a set of learnable filters (or kernels) to extract features like edges, textures, and shapes. A convolution operation, also referred to as a “convolution,” is used to perform many common image processing operations in the convolutional layer, by way of mathematical operations or algorithms. These operations include sharpening, blurring, noise reduction, embossing, and edge enhancement.

[0115] An activation function is a building block of neural networks that enable them to learn complex patterns in data. Activation functions transform the input signal of a node in the neural network into an output signal that is then passed on to the next layer. The node is where the “input signal,” i.e., the values of the input features, are summed after being multiplied by their corresponding weights and biases. At this stage, the activation function is applied. It transforms the summed weighted input into an output value, which is then passed on to the next neuron or group of neurons, i.e., the following layer. In effect, the activation layer activates the nodes, processes the input, generates an output proportional to the strength of the received signal, evaluates whether the output exceeds the threshold, and if so, passes it on to the neurons of the subsequent layer.

[0116] A rectified linear activation unit (ReLU) is a type of activation function. The ReLU function is a piecewise linear function that outputs the input directly if it is positive. Otherwise, the ReLu outputs zero. A Leaky ReLU is a type of activation function that uses a slight non-negative gradient when the input is negative. Another type of activation function is a parametric ReLU (PReLU). This type of activation function generalizes the traditional rectified unit with a slope for negative values. A concatenated ReLU (CReLU) is a type of activation function which preserves both positive and negative phase information while enforcing non-saturated non-linearity.

[0117] Quantization is a technique used to reduce the precision of a set of values, and in the context of video compression, it refers to reducing the precision of coefficients, leading to compression. A quantization parameter (QP) is a dimensionless value used in quantization during video coding. In particular, QP is a parameter that controls the amount of compression applied to an image or video frame. Higher QP values mean more compression and lower quality, while lower QP values mean less compression and higher quality. A quantization step (QS) is a value used in image and video coding processing to divide the transform coefficient during the encoding process. QP determines the QS size applied to transform coefficients within video frames.

[0118] In video coding, an image or video frame can be divided into smaller regions called slices. Each slice can have its own QP value (QPslice), allowing for different levels of compression across different parts of the frame. A base QP (QPbase) is a global QP value that is used as a starting point for the QP values of all slices. A neural network can be trained to predict the optimal slice QP values based on the base QP and the characteristics of the video content.

[0119] An enhancement is a transform processing technology that aims to improve the quality of a quantized transform coefficient by recovering the quantization loss in the transform coefficientquantization process, thereby enhancing the overall transform processing. This transform processing technology may be, for example, an algorithm that is applied to the coefficients used to generate the reconstructed image.

[0120] Low-rank decomposition, also known as matrix factorization, is a technique that approximates a matrix by a product of two smaller matrices. For example, this technique splits a large matrix into a product of two smaller matrices, thereby providing a means for compression that reduces the parameters of a model while preserving one or more topological properties.

[0121] The following acronyms are used herein, Coding Tree Block (CTB), Coding Tree Unit (CTU), Coding Unit (CU), Joint Video Experts Team (JVET), Picture Order Count (POC), Versatile Video Coding (WC), Working Draft (WD), High Efficiency Video Coding (HEVC), Convolutional Neural Network (CNN), Non-Separable Transform (NSST), Karhunen-Loeve transform (KLT), High Operating Point (HOP), I-frame, P-frame, B-frame (IPB), Rectified Linear Unit (ReLU), a Parametric Rectified Linear Unit (PReLU), Concatenated Rectified Linear Unit (CReLU), Quantization Parameter (QP), and Quantization Step (QS).

[0122] FIG. 1 is a flowchart of an example operating method 100 of coding a video signal. Specifically, a video signal is encoded at an encoder. The encoding process compresses the video signal by employing various mechanisms to reduce the video file size. A smaller file size allows the compressed video file to be transmitted toward a user, while reducing associated bandwidth overhead. The decoder then decodes the compressed video file to reconstruct the original video signal for display to an end user. The decoding process generally mirrors the encoding process to allow the decoder to consistently reconstruct the video signal.

[0123] At step 101, the video signal is input into the encoder. For example, the video signal may be an uncompressed video file stored in memory. As another example, the video file may be captured by a video capture device, such as a video camera, and encoded to support live streaming of the video. The video file may include both an audio component and a video component. The video component contains a series of image frames that, when viewed in a sequence, gives the visual impression of motion. The frames contain pixels that are expressed in terms of light, referred to herein as luma components (or luma samples), and color, which is referred to as chroma components (or color samples). In some examples, the frames may also contain depth values to support three- dimensional viewing.

[0124] At step 103, the video is partitioned into blocks. Partitioning includes subdividing the pixels in each frame into square and / or rectangular blocks for compression. For example, in High Efficiency Video Coding (HEVC) (also known as H.265 and MPEG-H Part 2) the frame can first be divided into coding tree units (CTUs), which are blocks of a predefined size (e.g., sixty-four pixels by sixty-four pixels). The CTUs contain both luma and chroma samples. Coding trees may be employed to divide the CTUs into blocks and then recursively subdivide the blocks until configurations are achieved that support further encoding. For example, luma components of a frame may be subdivided until the individual blocks contain relatively homogenous lighting values. Further, chroma components of a frame may be subdivided until the individual blocks contain relatively homogenous color values. Accordingly, partitioning mechanisms vary depending on the content of the video frames.

[0125] At step 105, various compression mechanisms are employed to compress the image blocks partitioned at step 103. For example, inter-prediction and / or intra-prediction may be employed. Inter-prediction is designed to take advantage of the fact that objects in a common scene tend to appear in successive frames. Accordingly, a block depicting an object in a reference frame need not be repeatedly described in adjacent frames. Specifically, an object, such as a table, may remain in a constant position over multiple frames. Hence the table is described once and adjacent frames can refer back to the reference frame. Pattern matching mechanisms may be employed to match objects over multiple frames. Further, moving objects may be represented across multiple frames, for example due to object movement or camera movement. As a particular example, a video may show an automobile that moves across the screen over multiple frames. Motion vectors can be employed to describe such movement. A motion vector is a two-dimensional vector that provides an offset from the coordinates of an object in a frame to the coordinates of the object in a reference frame. As such, inter-prediction can encode an image block in a current frame as a set of motion vectors indicating an offset from a corresponding block in a reference frame.

[0126] Intra-prediction encodes blocks in a common frame. Intra-prediction takes advantage of the fact that luma and chroma components tend to cluster in a frame. For example, a patch of green in a portion of a tree tends to be positioned adjacent to similar patches of green. Intra-prediction employs multiple directional prediction modes (e.g., thirty-three in HEVC), a planar mode, and a direct current (DC) mode. The directional modes indicate that a current block is similar / the same as samples of a neighbor block in a corresponding direction. Planar mode indicates that a series ofblocks along a row / column (e.g., a plane) can be interpolated based on neighbor blocks at the edges of the row. Planar mode, in effect, indicates a smooth transition of light / color across a row / column by employing a relatively constant slope in changing values. DC mode is employed for boundary smoothing and indicates that a block is similar / the same as an average value associated with samples of all the neighbor blocks associated with the angular directions of the directional prediction modes. Accordingly, intra-prediction blocks can represent image blocks as various relational prediction mode values instead of the actual values. Further, inter-prediction blocks can represent image blocks as motion vector values instead of the actual values. In either case, the prediction blocks may not exactly represent the image blocks in some cases. Any differences are stored in residual blocks. Transforms may be applied to the residual blocks to further compress the file.

[0127] At step 107, various filtering techniques may be applied. In HEVC, the filters are applied according to an in-loop filtering scheme. The block based prediction discussed above may result in the creation of blocky images at the decoder. Further, the block based prediction scheme may encode a block and then reconstruct the encoded block for later use as a reference block. The in-loop filtering scheme iteratively applies noise suppression filters, de-blocking filters, adaptive loop filters, and sample adaptive offset (SAO) filters to the blocks / frames. These filters mitigate such blocking artifacts so that the encoded file can be accurately reconstructed. Further, these filters mitigate artifacts in the reconstructed reference blocks so that artifacts are less likely to create additional artifacts in subsequent blocks that are encoded based on the reconstructed reference blocks.

[0128] Once the video signal has been partitioned, compressed, and filtered, the resulting data is encoded in a bitstream at step 109. The bitstream includes the data discussed above as well as any signaling data desired to support proper video signal reconstruction at the decoder. For example, such data may include partition data, prediction data, residual blocks, and various flags providing coding instructions to the decoder. The bitstream may be stored in memory for transmission toward a decoder upon request. The bitstream may also be broadcast and / or multicast toward a plurality of decoders. The creation of the bitstream is an iterative process. Accordingly, steps 101, 103, 105, 107, and 109 may occur continuously and / or simultaneously over many frames and blocks. The order shown in FIG. 1 is presented for clarity and ease of discussion, and is not intended to limit the video coding process to a particular order.

[0129] The decoder receives the bitstream and begins the decoding process at step 111. Specifically, the decoder employs an entropy decoding scheme to convert the bitstream intocorresponding syntax and video data. The decoder employs the syntax data from the bitstream to determine the partitions for the frames at step 111. The partitioning should match the results of block partitioning at step 103. Entropy encoding / decoding as employed in step 111 is now described. The encoder makes many choices during the compression process, such as selecting block partitioning schemes from several possible choices based on the spatial positioning of values in the input image(s). Signaling the exact choices may employ a large number of bins. As used herein, a bin is a binary value that is treated as a variable (e.g., a bit value that may vary depending on context). Entropy coding allows the encoder to discard any options that are clearly not viable for a particular case, leaving a set of allowable options. Each allowable option is then assigned a code word. The length of the code words is based on the number of allowable options (e.g., one bin for two options, two bins for three to four options, etc.) The encoder then encodes the code word for the selected option. This scheme reduces the size of the code words as the code words are as big as desired to uniquely indicate a selection from a small sub-set of allowable options as opposed to uniquely indicating the selection from a potentially large set of all possible options. The decoder then decodes the selection by determining the set of allowable options in a similar manner to the encoder. By determining the set of allowable options, the decoder can read the code word and determine the selection made by the encoder.

[0130] At step 113, the decoder performs block decoding. Specifically, the decoder employs reverse transforms to generate residual blocks. Then the decoder employs the residual blocks and corresponding prediction blocks to reconstruct the image blocks according to the partitioning. The prediction blocks may include both intra-prediction blocks and inter-prediction blocks as generated at the encoder at step 105. The reconstructed image blocks are then positioned into frames of a reconstructed video signal according to the partitioning data determined at step 111. Syntax for step 113 may also be signaled in the bitstream via entropy coding as discussed above.

[0131] At step 115, filtering is performed on the frames of the reconstructed video signal in a manner similar to step 107 at the encoder. For example, noise suppression filters, de-blocking filters, adaptive loop filters, and SAO filters may be applied to the frames to remove blocking artifacts. Once the frames are filtered, the video signal can be output to a display at step 117 for viewing by an end user.

[0132] FIG. 2 is a schematic diagram of an example coding and decoding (codec) system 200 for video coding. Specifically, codec system 200 provides functionality to support theimplementation of operating method 100. Codec system 200 is generalized to depict components employed in both an encoder and a decoder. Codec system 200 receives and partitions a video signal as discussed with respect to steps 101 and 103 in operating method 100, which results in a partitioned video signal 201. Codec system 200 then compresses the partitioned video signal 201 into a coded bitstream when acting as an encoder as discussed with respect to steps 105, 107, and 109 in method 100. When acting as a decoder, codec system 200 generates an output video signal from the bitstream as discussed with respect to steps 111, 113, 115, and 117 in operating method 100. The codec system 200 includes a general coder control component 211, a transform scaling and quantization component 213, an intra-picture estimation component 215, an intra-picture prediction component 217, a motion compensation component 219, a motion estimation component 221, a scaling and inverse transform component 229, a filter control analysis component 227, an in-loop filters component 225, a decoded picture buffer component 223, and a header formatting and context adaptive binary arithmetic coding (CAB AC) component 231. Such components are coupled as shown. In FIG. 2, black lines indicate movement of data to be encoded / decoded while dashed lines indicate movement of control data that controls the operation of other components. The components of codec system 200 may all be present in the encoder. The decoder may include a subset of the components of codec system 200. For example, the decoder may include the intra-picture prediction component 217, the motion compensation component 219, the scaling and inverse transform component 229, the in-loop filters component 225, and the decoded picture buffer component 223. These components are now described.

[0133] The partitioned video signal 201 is a captured video sequence that has been partitioned into blocks of pixels by a coding tree. A coding tree employs various split modes to subdivide a block of pixels into smaller blocks of pixels. These blocks can then be further subdivided into smaller blocks. The blocks may be referred to as nodes on the coding tree. Larger parent nodes are split into smaller child nodes. The number of times a node is subdivided is referred to as the depth of the node / coding tree. The divided blocks can be included in coding units (CUs) in some cases. For example, a CU can be a sub-portion of a CTU that contains a luma block, red difference chroma (Cr) block(s), and a blue difference chroma (Cb) block(s) along with corresponding syntax instructions for the CU. The split modes may include a binary tree (BT), triple tree (TT), and a quad tree (QT) employed to partition a node into two, three, or four child nodes, respectively, of varying shapes depending on the split modes employed. The partitioned video signal 201 is forwarded tothe general coder control component 211, the transform scaling and quantization component 213, the intra-picture estimation component 215, the filter control analysis component 227, and the motion estimation component 221 for compression.

[0134] The general coder control component 211 is configured to make decisions related to coding of the images of the video sequence into the bitstream according to application constraints. For example, the general coder control component 211 manages optimization of bitrate / bitstream size versus reconstruction quality. Such decisions may be made based on storage space / bandwidth availability and image resolution requests. The general coder control component 211 also manages buffer utilization in light of transmission speed to mitigate buffer underrun and overrun issues. To manage these issues, the general coder control component 211 manages partitioning, prediction, and filtering by the other components. For example, the general coder control component 211 may dynamically increase compression complexity to increase resolution and increase bandwidth usage or decrease compression complexity to decrease resolution and bandwidth usage. Hence, the general coder control component 211 controls the other components of codec system 200 to balance video signal reconstruction quality with bit rate concerns. The general coder control component 211 creates control data, which controls the operation of the other components. The control data is also forwarded to the header formatting and CAB AC component 231 to be encoded in the bitstream to signal parameters for decoding at the decoder.

[0135] The partitioned video signal 201 is also sent to the motion estimation component 221 and the motion compensation component 219 for inter-prediction. A frame or slice of the partitioned video signal 201 may be divided into multiple video blocks. Motion estimation component 221 and the motion compensation component 219 perform inter-predictive coding of the received video block relative to one or more blocks in one or more reference frames to provide temporal prediction. Codec system 200 may perform multiple coding passes, e.g., to select an appropriate coding mode for each block of video data.

[0136] Motion estimation component 221 and motion compensation component 219 may be highly integrated, but are illustrated separately for conceptual purposes. Motion estimation, performed by motion estimation component 221, is the process of generating motion vectors, which estimate motion for video blocks. A motion vector, for example, may indicate the displacement of a coded object relative to a predictive block. A predictive block is a block that is found to closely match the block to be coded, in terms of pixel difference. A predictive block may also be referred toas a reference block. Such pixel difference may be determined by sum of absolute difference (SAD), sum of square difference (SSD), or other difference metrics. HEVC employs several coded objects including a CTU, coding tree blocks (CTBs), and CUs. For example, a CTU can be divided into CTBs, which can then be divided into CBs for inclusion in CUs. A CU can be encoded as a prediction unit (PU) containing prediction data and / or a transform unit (TU) containing transformed residual data for the CU. The motion estimation component 221 generates motion vectors, PUs, and TUs by using a rate-distortion analysis as part of a rate distortion optimization process. For example, the motion estimation component 221 may determine multiple reference blocks, multiple motion vectors, etc. for a current block / frame, and may select the reference blocks, motion vectors, etc. having the best rate-distortion characteristics. The best rate-distortion characteristics balance both quality of video reconstruction (e.g., amount of data loss by compression) with coding efficiency (e.g., size of the final encoding).

[0137] In some examples, codec system 200 may calculate values for sub-integer pixel positions of reference pictures stored in decoded picture buffer component 223. For example, video codec system 200 may interpolate values of one-quarter pixel positions, one-eighth pixel positions, or other fractional pixel positions of the reference picture. Therefore, motion estimation component 221 may perform a motion search relative to the full pixel positions and fractional pixel positions and output a motion vector with fractional pixel precision. The motion estimation component 221 calculates a motion vector for a PU of a video block in an inter-coded slice by comparing the position of the PU to the position of a predictive block of a reference picture. Motion estimation component 221 outputs the calculated motion vector as motion data to header formatting and CAB AC component 231 for encoding and motion to the motion compensation component 219.

[0138] Motion compensation, performed by motion compensation component 219, may involve fetching or generating the predictive block based on the motion vector determined by motion estimation component 221. Again, motion estimation component 221 and motion compensation component 219 may be functionally integrated, in some examples. Upon receiving the motion vector for the PU of the current video block, motion compensation component 219 may locate the predictive block to which the motion vector points. A residual video block is then formed by subtracting pixel values of the predictive block from the pixel values of the current video block being coded, forming pixel difference values. In general, motion estimation component 221 performs motion estimation relative to luma components, and motion compensation component 219 uses motion vectorscalculated based on the luma components for both chroma components and luma components. The predictive block and residual block are forwarded to transform scaling and quantization component 213.

[0139] The partitioned video signal 201 is also sent to intra-picture estimation component 215 and intra-picture prediction component 217. As with motion estimation component 221 and motion compensation component 219, intra-picture estimation component 215 and intra-picture prediction component 217 may be highly integrated, but are illustrated separately for conceptual purposes. The intra-picture estimation component 215 and intra-picture prediction component 217 intra-predict a current block relative to blocks in a current frame, as an alternative to the inter-prediction performed by motion estimation component 221 and motion compensation component 219 between frames, as described above. In particular, the intra-picture estimation component 215 determines an intraprediction mode to use to encode a current block. In some examples, intra-picture estimation component 215 selects an appropriate intra- prediction mode to encode a current block from multiple tested intra-prediction modes. The selected intra-prediction modes are then forwarded to the header formatting and CAB AC component 231 for encoding.

[0140] For example, the intra-picture estimation component 215 calculates rate-distortion values using a rate-distortion analysis for the various tested intra-prediction modes, and selects the intraprediction mode having the best rate-distortion characteristics among the tested modes. Ratedistortion analysis generally determines an amount of distortion (or error) between an encoded block and an original unencoded block that was encoded to produce the encoded block, as well as a bitrate (e.g., a number of bits) used to produce the encoded block. The intra-picture estimation component 215 calculates ratios from the distortions and rates for the various encoded blocks to determine which intra-prediction mode exhibits the best rate-distortion value for the block. In addition, intra-picture estimation component 215 may be configured to code depth blocks of a depth map using a depth modeling mode (DMM) based on rate-distortion optimization (RDO).

[0141] The intra-picture prediction component 217 may generate a residual block from the predictive block based on the selected intra-prediction modes determined by intra-picture estimation component 215 when implemented on an encoder or read the residual block from the bitstream when implemented on a decoder. The residual block includes the difference in values between the predictive block and the original block, represented as a matrix. The residual block is then forwarded to the transform scaling and quantization component 213. The intra-picture estimation component215 and the intra-picture prediction component 217 may operate on both luma and chroma components.

[0142] The transform scaling and quantization component 213 is configured to further compress the residual block. The transform scaling and quantization component 213 applies a transform, such as a discrete cosine transform (DCT), a discrete sine transform (DST), or a conceptually similar transform, to the residual block, producing a video block comprising residual transform coefficient values. Wavelet transforms, integer transforms, sub-band transforms or other types of transforms could also be used. The transform may convert the residual information from a pixel value domain to a transform domain, such as a frequency domain. The transform scaling and quantization component 213 is also configured to scale the transformed residual information, for example based on frequency. Such scaling involves applying a scale factor to the residual information so that different frequency information is quantized at different granularities, which may affect final visual quality of the reconstructed video. The transform scaling and quantization component 213 is also configured to quantize the transform coefficients to further reduce bit rate. The quantization process may reduce the bit depth associated with some or all of the coefficients. The degree of quantization may be modified by adjusting a quantization parameter. In some examples, the transform scaling and quantization component 213 may then perform a scan of the matrix including the quantized transform coefficients. The quantized transform coefficients are forwarded to the header formatting and CAB AC component 231 to be encoded in the bitstream.

[0143] The scaling and inverse transform component 229 applies a reverse operation of the transform scaling and quantization component 213 to support motion estimation. The scaling and inverse transform component 229 applies inverse scaling, transformation, and / or quantization to reconstruct the residual block in the pixel domain, e.g., for later use as a reference block which may become a predictive block for another current block. The motion estimation component 221 and / or motion compensation component 219 may calculate a reference block by adding the residual block back to a corresponding predictive block for use in motion estimation of a later block / frame. Filters are applied to the reconstructed reference blocks to mitigate artifacts created during scaling, quantization, and transform. Such artifacts could otherwise cause inaccurate prediction (and create additional artifacts) when subsequent blocks are predicted.

[0144] The filter control analysis component 227 and the in-loop filters component 225 apply the filters to the residual blocks and / or to reconstructed image blocks. For example, the transformedresidual block from the scaling and inverse transform component 229 may be combined with a corresponding prediction block from intra-picture prediction component 217 and / or motion compensation component 219 to reconstruct the original image block. The fdters may then be applied to the reconstructed image block. In some examples, the fdters may instead be applied to the residual blocks. As with other components in FIG. 2, the fdter control analysis component 227 and the in-loop fdters component 225 are highly integrated and may be implemented together, but are depicted separately for conceptual purposes. Filters applied to the reconstructed reference blocks are applied to particular spatial regions and include multiple parameters to adjust how such fdters are applied. The fdter control analysis component 227 analyzes the reconstructed reference blocks to determine where such fdters should be applied and sets corresponding parameters. Such data is forwarded to the header formatting and CAB AC component 231 as fdter control data for encoding. The in-loop fdters component 225 applies such fdters based on the fdter control data. The fdters may include a deblocking fdter, a noise suppression fdter, a SAG fdter, and an adaptive loop fdter. Such fdters may be applied in the spatial / pixel domain (e.g., on a reconstructed pixel block) or in the frequency domain, depending on the example.

[0145] When operating as an encoder, the fdtered reconstructed image block, residual block, and / or prediction block are stored in the decoded picture buffer component 223 for later use in motion estimation as discussed above. When operating as a decoder, the decoded picture buffer component 223 stores and forwards the reconstructed and filtered blocks toward a display as part of an output video signal. The decoded picture buffer component 223 may be any memory device capable of storing prediction blocks, residual blocks, and / or reconstructed image blocks.

[0146] The header formatting and CAB AC component 231 receives the data from the various components of codec system 200 and encodes such data into a coded bitstream for transmission toward a decoder. Specifically, the header formatting and CAB AC component 231 generates various headers to encode control data, such as general control data and filter control data. Further, prediction data, including intra-prediction and motion data, as well as residual data in the form of quantized transform coefficient data are all encoded in the bitstream. The final bitstream includes all information desired by the decoder to reconstruct the original partitioned video signal 201. Such information may also include intra-prediction mode index tables (also referred to as codeword mapping tables), definitions of encoding contexts for various blocks, indications of most probable intra-prediction modes, an indication of partition information, etc. Such data may be encoded byemploying entropy coding. For example, the information may be encoded by employing context adaptive variable length coding (CAVLC), CABAC, syntax-based context-adaptive binary arithmetic coding (SB AC), probability interval partitioning entropy (PIPE) coding, or another entropy coding technique. Following the entropy coding, the coded bitstream may be transmitted to another device (e.g., a video decoder) or archived for later transmission or retrieval.

[0147] FIG. 3 is a block diagram illustrating an example video encoder 300. Video encoder 300 may be employed to implement the encoding functions of codec system 200 and / or implement steps 101, 103, 105, 107, and / or 109 of operating method 100. Encoder 300 partitions an input video signal, resulting in a partitioned video signal 301, which is substantially similar to the partitioned video signal 201. The partitioned video signal 301 is then compressed and encoded into a bitstream by components of encoder 300.

[0148] Specifically, the partitioned video signal 301 is forwarded to an intra-picture prediction component 317 for intra-prediction. The intra-picture prediction component 317 may be substantially similar to intra-picture estimation component 215 and intra-picture prediction component 217. The partitioned video signal 301 is also forwarded to a motion compensation component 321 for inter-prediction based on reference blocks in a decoded picture buffer component 323. The motion compensation component 321 may be substantially similar to motion estimation component 221 and motion compensation component 219. The prediction blocks and residual blocks from the intra-picture prediction component 317 and the motion compensation component 321 are forwarded to a transform and quantization component 313 for transform and quantization of the residual blocks. The transform and quantization component 313 may be substantially similar to the transform scaling and quantization component 213. The transformed and quantized residual blocks and the corresponding prediction blocks (along with associated control data) are forwarded to an entropy coding component 331 for coding into a bitstream. The entropy coding component 331 may be substantially similar to the header formatting and CABAC component 231.

[0149] The transformed and quantized residual blocks and / or the corresponding prediction blocks are also forwarded from the transform and quantization component 313 to an inverse transform and quantization component 329 for reconstruction into reference blocks for use by the motion compensation component 321. The inverse transform and quantization component 329 may be substantially similar to the scaling and inverse transform component 229. In-loop fdters in an inloop fdters component 325 are also applied to the residual blocks and / or reconstructed referenceblocks, depending on the example. The in-loop fdters component 325 may be substantially similar to the filter control analysis component 227 and the in-loop filters component 225. The in-loop filters component 325 may include multiple filters as discussed with respect to in-loop filters component 225. The filtered blocks are then stored in a decoded picture buffer component 323 for use as reference blocks by the motion compensation component 321. The decoded picture buffer component 323 may be substantially similar to the decoded picture buffer component 223.

[0150] FIG. 4 is a block diagram illustrating an example video decoder 400. Video decoder 400 may be employed to implement the decoding functions of codec system 200 and / or implement steps 111, 113, 115, and / or 117 of operating method 100. Decoder 400 receives a bitstream, for example from an encoder 300, and generates a reconstructed output video signal based on the bitstream for display to an end user.

[0151] The bitstream is received by an entropy decoding component 433. The entropy decoding component 433 is configured to implement an entropy decoding scheme, such as CAVLC, CAB AC, SBAC, PIPE coding, or other entropy coding techniques. For example, the entropy decoding component 433 may employ header information to provide a context to interpret additional data encoded as codewords in the bitstream. The decoded information includes any desired information to decode the video signal, such as general control data, filter control data, partition information, motion data, prediction data, and quantized transform coefficients from residual blocks. The quantized transform coefficients are forwarded to an inverse transform and quantization component 429 for reconstruction into residual blocks. The inverse transform and quantization component 429 may be similar to inverse transform and quantization component 329.

[0152] The reconstructed residual blocks and / or prediction blocks are forwarded to intra-picture prediction component 417 for reconstruction into image blocks based on intra-prediction operations. The intra-picture prediction component 417 may be similar to intra-picture estimation component 215 and to intra-picture prediction component 217. Specifically, the intra-picture prediction component 417 employs prediction modes to locate a reference block in the frame and applies a residual block to the result to reconstruct intra-predicted image blocks. The reconstructed intrapredicted image blocks and / or the residual blocks and corresponding inter-prediction data are forwarded to a decoded picture buffer component 423 via an in-loop filters component 425, which may be substantially similar to decoded picture buffer component 223 and in-loop filters component 225, respectively. The in-loop filters component 425 filters the reconstructed image blocks, residualblocks and / or prediction blocks, and such information is stored in the decoded picture buffer component 423. Reconstructed image blocks from decoded picture buffer component 423 are forwarded to a motion compensation component 421 for inter-prediction. The motion compensation component 421 may be substantially similar to motion estimation component 221 and / or motion compensation component 219. Specifically, the motion compensation component 421 employs motion vectors from a reference block to generate a prediction block and applies a residual block to the result to reconstruct an image block. The resulting reconstructed blocks may also be forwarded via the in-loop filters component 425 to the decoded picture buffer component 423. The decoded picture buffer component 423 continues to store additional reconstructed image blocks, which can be reconstructed into frames via the partition information. Such frames may also be placed in a sequence. The sequence is output toward a display as a reconstructed output video signal.

[0153] Keeping the above in mind, video compression techniques perform spatial (intra-picture) prediction and / or temporal (inter-picture) prediction to reduce or remove redundancy inherent in video sequences. For block-based video coding, a video slice (i.e., a video picture or a portion of a video picture) may be partitioned into video blocks, which may also be referred to as treeblocks, coding tree blocks (CTBs), coding tree units (CTUs), coding units (CUs), and / or coding nodes. Video blocks in an intra-coded (I) slice of a picture are encoded using spatial prediction with respect to reference samples in neighboring blocks in the same picture. Video blocks in an inter-coded (P or B) slice of a picture may use spatial prediction with respect to reference samples in neighboring blocks in the same picture or temporal prediction with respect to reference samples in other reference pictures. Pictures may be referred to as frames, and reference pictures may be referred to as reference frames. The POC is a variable associated with each picture that uniquely identifies the associated picture among all pictures in the coded layer video sequence (CLVS), indicates when the associated picture is to be output from the decoded picture buffer (DPB), and indicates the position of the associated picture in output order relative to the output order positions of the other pictures in the same CLVS that are to be output from the DPB. A flag is a variable or single-bit syntax element that can take one of the two possible values: 0 and 1.

[0154] Spatial or temporal prediction results in a predictive block for a block to be coded. Residual data represents pixel differences between the original block to be coded and the predictive block. An inter-coded block is encoded according to a motion vector that points to a block of reference samples forming the predictive block, and the residual data indicating the differencebetween the coded block and the predictive block. An intra-coded block is encoded according to an intra-coding mode and the residual data. For further compression, the residual data may be transformed from the pixel domain to a transform domain, resulting in residual transform coefficients, which then may be quantized. The quantized transform coefficients, initially arranged in a two-dimensional array, may be scanned in order to produce a one-dimensional vector of transform coefficients, and entropy coding may be applied to achieve even more compression.

[0155] Image and video compression have experienced rapid growth, leading to various coding standards. Such video coding standards include ITU-T H.261, International Organization for Standardization / International Electrotechnical Commission (ISO / IEC) MPEG-1 Part 2, ITU-T H.262 or ISO / IEC MPEG-2 Part 2, ITU-T H.263, ISO / IEC MPEG-4 Part 2, Advanced Video Coding (AVC), also known as ITU-T H.264 or ISO / IEC MPEG-4 Part 10, and High Efficiency Video Coding (HEVC), also known as ITU-T H.265 or MPEG-H Part 2. AVC includes extensions such as Scalable Video Coding (SVC), Multiview Video Coding (MVC) and Multiview Video Coding plus Depth (MVC+D), and 3D AVC (3D-AVC). HEVC includes extensions such as Scalable HEVC (SHVC), Multiview HEVC (MV-HEVC), and 3D HEVC (3D-HEVC).

[0156] Versatile Video Coding (WC), also known as H.266, ISO / IEC 23090-3, and MPEG-I Part 3, was finalized on July 6, 2020, by the Joint Video Experts Team (JVET) of the VCEG working group of ITU-T Study Group 16 and the MPEG working group of ISO / IEC JTC 1 / SC 29. The description of the techniques disclosed herein are based on the WC by the JVET of ITU-T and ISO / IEC. However, the techniques also apply to other video codec specifications.

[0157] Current neural network-based video compression tools face challenges, particularity with respect to reconstruction accuracy of inverse transforms in the video coding process. The model architecture is not fully optimized, resulting in inefficient complexity reduction or performance improvement. With the rapid progress of deep learning research, the development of neural networkbased video compression tools shows very promising results by utilizing many state-of-the-art deep learning techniques. These tools enhance various modules within traditional video coding algorithms. For instance, convolutional neural networks assisted intra prediction and in-loop filters in WC yield promising outcomes in terms of image coding efficiency.

[0158] The convolutional process involves moving filters across images, applying them horizontally and vertically with specified strides. Residual blocks, including convolutional layers and activation functions, are frequently employed in deep convolutional neural network (CNN)architectures. During inference, distorted reconstruction frames are input into pre-trained CNN models. These models process frames either before or after certain compression stages (e.g., deblocking, sample adaptive offset, or adaptive in-loop filter).

[0159] Transform coding is a component of video codecs used for achieving high compression ratios. Traditional DCT, particularly DCT-2, has been a foundation of transform coding in the past few decades because of its reasonable balance between coding performance and complexity. DCT- 2 is mathematically proven to approximate the optimal data-driven Karhunen-Loeve transform (KLT) under the first-order Markov conditions, which efficiently model natural imagery sources.

[0160] In H.264 / AVC, a low-complexity 4x4 transform is utilized, featuring multiplier-less calculation, fixed-point, and 16-bit intermediate data representation.

[0161] HEVC improves on transform coding by extending the DCT-2 transform sizes to 8x8, 16x16, and 32x32 with fixed-point kernels. The transform implementation can be done using either direct matrix multiplication or a partial butterfly fast method. The design also maintains a 16-bit intermediate data representation and arithmetic. Additionally, HEVC uses a 4x4 DST-7 for 4x4 intra prediction residuals.

[0162] In VVC, transform signalling has been adopted to overcome the limitation of using a single fixed kernel. This allows multiple options of transform kernels, so an encoder can choose a transform on a per-block basis and signal such a selection in the bitstream. Transform candidates can be derived by off-line training or selected from a group of mathematically defined transforms, such as the DCT / DST families. A secondary transform is an additional transform process that follows the primary transform. The major benefit of applying a non-separable transform (NSST) as a secondary transform is to achieve a better trade-off between coding efficiency and complexity. With NSST, a non-separable secondary transform is performed on the lower-frequency coefficients to largely reduce computational complexity.

[0163] As we can see from the coding loop, quantization is performed after transformation at encoder side, and dequantization is performed prior to the inverse transformation at decoder side. Quantization error is propagated through inverse transformation which lead to a suboptimal reconstruction result.(1) Inverse Transform Enhancement

[0164] Disclosed herein are techniques that improve reconstruction accuracy of inverse transforms by applying an enhancement process to inverse transform coefficients. In particular, acoefficient enhancement is applied to transform coefficients before applying the inverse transform. By applying the enhancement before the transform, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0165] FIG. 5 illustrates an embodiment of a coding method 500 implemented by an encoding device in a CNN. At step 505, the encoding device applies an enhancement to transform coefficients. An enhancement is a transform processing technology that aims to improve the quality of a quantized transform coefficient by recovering the quantization loss in the transform coefficient quantization process, thereby enhancing the overall transform processing. This transform processing technology may be, for example, an algorithm that is applied to the coefficients used to generate the reconstructed image. At step 510, the encoding device generates enhanced transform coefficients based on the enhancement applied to the transform coefficients. At step 515, the encoding device applies an inverse transform to the enhanced transform coefficients to generate a reconstructed transform. In this way, the encoding device applies the enhancement before applying the inverse transform operation to generate the reconstructeed transform. At step 520, the encoding dervice generates a bitstream based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer- readable medium.

[0166] FIG. 6 illustrates an embodiment of a coding method 600 implemented by a decoding device in a CNN. At step 605, the decoding device applies an enhancement to transform coefficients. At step 610, the decoding device generates enhanced transform coefficients generated based on the enhancement. At step 615, the decoding device applies an inverse transform to the enhanced transform coefficients to generate a reconstructed transform. At step 620, the decoding device generates an image based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer- readable medium.

[0167] FIG. 7 illustrates an embodiment of an inverse transform enhancement 700 as performed, for example, in the scaling and inverse transform component 229 in FIG. 2. The enhancement 700may be implemented by a coding device, such as the encoder 300, as illustrated in FIG. 3, or the decoder 400, as illustrated in FIG. 4. The enhancement 700 is a process, for example, the coding method 500 as implemented by an encoding device or the coding method 600 as implemented by a decoding device, to enhance inverse transformations in the video coding process.

[0168] As shown in FIG. 7, the scaling and inverse transform component 229 receives, as input, transform coefficients from the transform scaling and quantization component 213 of FIG. 2. Specifically, an inverse secondary transform enhancement component 900, which includes a coefficient enhancement component 710 and an inverse secondary transform component 715, receives the transform coefficients at the coefficient enhancement component 710. The coefficient enhancement component 710 applies an enhancement to the transform coefficients, and based on the enhancement, generates enhanced transform coefficients, which are input to the inverse secondary transform component 715. The inverse secondary transform component 715 applies an inverse transform to the enhanced transform coefficients, and generates a reconstructed transform based on the enhanced transform coefficients. A bitstream or an image may be generated based on the reconstructed transform generated by the inverse secondary transform 715.

[0169] The enhancement 700 may further include an inverse primary transform enhancement component 800. The inverse primary transform enhancement component 800 may include a coefficient enhancement component 720, an inverse primary vertical transform component 725, a vertical enhancement component 730, an inverse primary horizontal transform component 735, and a horizontal enhancement component 740. As further illustrated, the inverse primary transform enhancement component 800 receives, as input, transform coefficients of the reconstructed transform from the inverse secondary transform enhancement 900. Specifically, the coefficient enhancement component 720 receives the transform coefficients, applies an enhancement to the transform coefficients, and generates transform coefficients, which are further enhanced based on the enhancement applied by the coefficient enhancement component 720. The transform coefficients are input into the inverse primary transform component 725, which applies an inverse transform to the enhanced transform coefficients, thereby generating a reconstructed transform. A bistream or an image may be generated based on the reconstructed transform.

[0170] As shown in the embodiment of FIG. 7, multiple enhancement processes may be provided to enhance the inverse transform. For example, the coefficient enhancement component 710 is applied before the inverse secondary transform component 715, and the coefficientenhancement component 720 is applied after the inverse secondary transform component 715, and before the inverse primary vertical transform component 725. Additionally, a vertical enhancement process 730 may be applied after the inverse primary vertical transform component 725, and before the inverse primary horizontal transform component 735. Even further, a horizontal enhancement component 740 may be used after the inverse primary horizontal transform component 735, where the horizontal enhancement component outputs a reconstructed transform. These components of the enhancement 700 may be dedicated for different transform types, or the transform types may be used as inputs to the components so that the components can be shared among different transforms.

[0171] FIG. 8 illustrates an embodiment of an inverse primary transform enhancement 800 implemented by a coding device. FIG. 8 is a more detailed illustration of the inverse primary transform enhancement 800 shown in FIG. 7. As shown, QS component 805 receives a QS as input instead of a QPbase or a QPslice. The QS component 805 processes the QS and outputs a latent tensor QS, which is input to a fuse component 825, which also receives input from a transform (TR) component 820. That is, TR component 820 performs processing on incoming transform coefficients and outputs a latent representation TR, which is input to the fuse component 825. The fuse component 825 combines the QS and TR, and outputs a latent tensor, which passes through one or more back bone blocks 830 for the extraction of more features. The extracted features from the back bone blocks 830 are input to a restoration (RESTO) module, which processes the output of last back bone block 830 and generates an enhanced transform coefficient.

[0172] The inverse vertical transform component 840 receives the enhanced transform coefficients as input, performs a vertical transform on the enhanced transform coefficients, and outputs the results of the vertical transform to a summation operation 845. The summation operation 845 also receives input from a vertical scaling component 810, which receives the latent tensor QS from the QS component 805. The vertical scaling component 810 performs vertical scaling on the latent tensor QS, and generates scaling and bias parameters as input to both a horizontal scaling component 815 and the summation operation 845, which combines the output of the inverse vertical transform 840 and the vertical scaling component 810. In this way, the output of the inverse vertical transform component 840 is adjusted based on the scaling and bias parameters in the summation operation 845.

[0173] The output from the summation operation 845 is input into the inverse horizontal transform component 850. The inverse horizontal transform component 850 performs an inversehorizontal transform on the adjusted output of the inverse vertical transform component 840 received from the summation operation 845. The horizontal scaling component 815, which receives scaling and bias parameters from the vertical scaling component 810, and receives the latent tensor QS from the QS component 805, performs horizontal scaling on the parameters received from the vertical scaling component 810 and the QS component 805. The horizontal scaling component 815 generates scaling and bias components for input into another summing operation 855. The output generated by the inverse horizontal transform component 850 is adjusted based on the scaling and bias parameters generated as output from the horizontal scaling component 815, where the output from the horizontal scaling component 815 is summed, in the summation operation 855, with the output of the inverse horizontal transform component 850. The output of the summation operation 855 is used to generate a reconstructed transform.

[0174] Various neural networks can be used as the module structure in the framework illustrated in FIG. 8. For example, in an embodiment, a QS component 805 and the TR component 820 contain sequential convolutional and rectified linear unit (Conv-ReLU) operations to convert an input signal from 1 channel to N channels. Similarly, in an embodiment, the fuse component 825 contains sequential Conv-ReLU-Conv operations to combine and consolidate multiple inputs into one output. Additionally, in an embodiment one or more back bone blocks 830 contain a residual block, which uses sequential Conv-ReLU-Conv-Conv operations followed by a skip summation operation to extract more features by expanding and restoring the channels. In an embodiment, a RESTO module 835 contains sequential Conv-PixelShuffle operations to adjust the feature tensor to the right order and shape. In an embodiment, inverse vertical / horizontal transform modules contain a linear layer to perform the inverse vertical / horizontal transform. In an embodiment, vertical / horizontal scaling modules contain two trainable tensors to perform element-wise scaling and bias summation.

[0175] In an embodiment, each pair of vertical and horizontal transform types uses its dedicated enhancement module.

[0176] In another embodiment, the vertical and horizontal transform types are used as inputs to an enhancement module, so that the enhancement module can be shared among different transform types.

[0177] In yet another embodiment, every pair of vertical and horizontal transforms is supplied with a dedicated or shared enhancement module.

[0178] In another embodiment, only a selected one or more pairs of vertical and horizontal transforms are supplied with dedicated or shared enhancement modules.

[0179] In an embodiment, the transform coefficients have a different size and shape according to the block partitioning decision from the encoding process, and are supplied with either a dedicated enhancement module, a fully shared enhancement module, or a partially shared enhancement module.

[0180] In an embodiment, CReLU replaces an activation function, such as ReLU or PReLU, in one or more layers: CReLU(x) = ReLU( concat( x, -x) ), where x is the output tensor of a previous layer. In this case, x and -x are concatenated and a concatenated tensor is generated and provided as the input to the ReLU function. If it is desired to keep the output channel size of ReLU operation unchanged, the output channel size of the previous layer is cut to half, resulting in decreasing the number of parameters for some or all previous layers.

[0181] FIG. 9 illustrates an embodiment of an inverse secondary transform enhancement 900 implemented by a coding device. FIG. 9 is a more detailed illustration of the inverse secondary transform enhancement 800 shown in FIG. 7. As shown in FIG. 9, a QS component 905 receives a QS as input instead of QPbase or QPslice, and upon processing the QS, the QS component 905 outputs a latent tensor QS. In a parallel operation, a TR component 915 processes an incoming transform coefficients and outputs a latent representation TR. A fuse component 920 combines the QS latent tensor and the TR latent representation, and outputs a latent tensor, which passes through one or more back bone blocks 925 for additional feature extraction. The extracted features from the back bone blocks 925 are input to a RESTO module 930, which processes the output of the last back bone block 925 and generates an enhanced transform coefficient.

[0182] The inverse vertical transform component 935 receives the enhanced transform coefficient as input, and applies an inverse secondary transform to the enhanced transform coefficient. A scaling component 910 receives the QS from the QS component 905, performs scaling of the QS, and generates scaling and bias parameters, which are summed with the output of the inverse transform component 935 in a summation operation 940. In this way, the output of the inverse transform component 935 is adjusted based on the scaling and bias parameters in the summation operation 940. The output of the summation operation 940 is used to generate a reconstructed transform.

[0183] Various neural networks can be used as the module structure in the framework illustrated in FIG. 9. For example, in an embodiment, a QS component 905 and the TR component 915 contain sequential convolutional and rectified linear unit (Conv-ReLU) operations to convert an input signal from 1 channel to N channels. Similarly, in an embodiment, the fuse component 920 contains sequential Conv-ReLU-Conv operations to combine and consolidate multiple inputs into one output. Additionally, in an embodiment, one or more back bone blocks 925 contain a residual block, which uses sequential Conv-ReLU-Conv-Conv operations followed by a skip summation operation to extract more features by expanding and restoring the channels. In an embodiment, the RESTO module 930 contains sequential Conv-PixelShuffle operations to adjust the feature tensor to the right order and shape. In an embodiment, inverse vertical / horizontal transform modules contain a linear layer to perform the inverse vertical / horizontal transform. In an embodiment, vertical / horizontal scaling modules contain two trainable tensors to perform element -wise scaling and bias summation.

[0184] In an embodiment, each pair of vertical and horizontal transform types uses a dedicated enhancement module.

[0185] In another embodiment, the vertical and horizontal transform types are used as inputs to an enhancement module, so that the enhancement module can be shared among different transform types.

[0186] In yet another embodiment, every pair of vertical and horizontal transforms is supplied with a dedicated or shared enhancement module.

[0187] In another embodiment, only a selected one or more pairs of vertical and horizontal transforms are supplied with dedicated or shared enhancement modules.

[0188] In an embodiment, the transform coefficients have a different size and shape according to the block partitioning decision from the encoding process, and are supplied with either a dedicated enhancement module, a fully shared enhancement module, or a partially shared enhancement module.

[0189] In an embodiment, CReLU is used to replace an activation function, such as ReLU or PReLU, in one or more layers: CReLU(x) = ReLU( concat( x, -x) ), where x is the output tensor of a previous layer. In this case, x and -x are concatenated and a concatenated tensor is generated as the input of a ReLU function. If it is desired to keep the output channel size of ReLU operation unchanged, the output channel size of the previous layer is cut to half, resulting in decreasing the number of parameters for some or all previous layers.(2) In-Loop Filter Control

[0190] Disclosed herein are techniques that improve reconstruction accuracy of inverse transforms by providing direct interpretation of input signals to control the quantization level. In particular, a QS is applied as input during a quantization process instead of applying QPs. By applying QS to control the quantization level, overly complex processing may be avoided during reconstruction of the inverse transforms. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improved video coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0191] FIG. 10 is an example of a neural network-based loop filter 1000, as implemented in the in-loop filters component 225 of FIG. 2. Specifically, the filter 1000 is an example of a high operating point (HOP) neural network-based loop filter. As illustrated, the filter 1000 includes a first layer 1005, a second layer 1010, and a third layer 1015. The first layer includes, as input parameters, an I-frame, P-frame, B-frame (IPB) indicator, a slice quantization parameter (QPslice), a base quantization parameter (QPbase), boundary strength (BS), a prediction mode (Pred), and a reconstructed image block (Rec). The second layer 1010 includes convolution layers Convlxl, Convlx3, Conv3xl, Conv3x3 as convolution layers. The third layer 1015 includes activation functions, wich may be a rectified linear unit (ReLU) or, as illustrated in FIG. 10, the activation functions may be a parametric rectified linear unit (PReLU) or a Leaky ReLU.

[0192] As further illustrated in FIG. 10, the output from the third layer 1015 is input into sequential Conv-PReLU-Conv-PReLU operations, and the output from the sequential Conv- PReLU-Conv-PReLU operations is input to one or more back bone blocks. The back bone blocks each contain a series of conv layers and a PReLU, and the output of the last back bone block is input into sequential Conv-PReLU-Conv operations. A pixel shuffle component receives, as input, the output from the last conv layer.

[0193] In the filter 1000, QPslice and QPbase, which may be collectively referred to as quantization parameters (QP) or “QPs” in this disclosure, are input to the second layer 1010, which is the convolution layer, where they are processed to obtain QS. In this regard, even though QPs are used to control the quantization level, QS represents the size of the quantization intervals and is the step size used in the quantization process. As a result of inputting the QP to the convolution layer,QPslice and QPbase are combined and converted to QS. This conversion involves a complex quantization process and involves specifics of the video codec’s quantization algorithm, including the quantization matrix, rate control mechanisms, and may further include various other factors that contribute to the complexity of the process.

[0194] In order to reduce the complexity resulting from the example shown in FIG. 10, and instead of using the parameters QPslice and QPbase as auxiliary input information for both training and inference of NN-based coding tools, an embodiment of this disclosure combines QPslice and QPbase to form the final QP used by the convolutional component processing block, then converts the QP to QS, which is input as auxiliary input information for both training and inference of NN- based coding tools. In this way, QS replaces QP (i.e., QPslice and QPbase) in the first layer 1005. This replacement provides the neural network with a more direct interpretation of input signals, thereby providing the neural network with more capacity to improve model performance.

[0195] FIG. HA illustrates an embodiment of layers 1100 applied in a neural network-based loop filter, where QS replaces QP. In particular, the layers 1100 include a first layer 1005a, a second layer 1010, and a third layer 1015. In the first layer 1005a, QS is directly input to the second layer 1010 to control the quantization level. More specifically, instead of inputting QPslice and QPbase in the first layer 1005 as shown in FIG. 10, QS is input to the first layer 1005a as shown in FIG. 11. As a result of the direct input of QS, the loop filter avoids performing complex computations involved with converting QP to QS in the second layer 1010. Thus, network capacity is increased and network performance is improved.

[0196] FIG. 11B also illustrates an embodiment of layers 1100 applied in a neural networkbased loop filter. In particular, the layers 1100 include a first layer 1005b, a second layer 1010, and a third layer 1015. In the first layer 1005a, QS is directly input to the second layer 1010 so as to control the quantization level by using a single QS component. More specifically, instead of inputting QPslice and QPbase in the first layer 1005 as shown in FIG. 10, QS is input to the first layer 1005b as shown in FIG. 11. Thus, network capacity is increased and network performance is improved.

[0197] The embodiments shown in FIG. 11 A and FIG. 1 IB are improvements over the example shown in FIG. 10. In this regard, instead of using the parameters QPslice and QPbase, as illustrated in FIG. 10, as auxiliary input information for both training and inference of NN-based coding tools, an embodiment of this disclosure combines QPslice and QPbase to form the final QP used by theconvolutional component processing block, then converts the QP to QS as auxiliary input information for both training and inference of NN-based coding tools. In this way, QS replaces QP (i.e., QPslice and QPbase) in the first layer 1005. This replacement provides the neural network with a direct interpretation of input signals and provides the neural network with more capacity to improve model performance.

[0198] FIG. 12 illustrates an embodiment of a quantization step applied in a coding method 1200 implemented by an encoding device. At step 1205, the encoding device receives transform coefficients. At step 1210, the encoding device applies a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform. At step 1215, the encoding device generates a bitstream based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.

[0199] FIG. 13 illustrates an embodiment of a quantization step applied in a coding method 1300 implemented by a decoding device. At step 1305, the decoding device receives transform coefficients. At step 1310, the decoding device applies a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform. At step 1315, the decoding device generates an image based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.

[0200] FIG. 14 illustrates an embodiment of a quantization step and inverse transform enhancement applied in a coding method 1400 implemented by an encoding device. At step 1405, the encoding device receives transform coefficients. At step 1410, the encoding device generates a reconstructed transform. The encoding device generates the reconstructed transform by: applying, at step 1415, a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP; applying, at step 1420, an enhancement to the transform coeeficients; and applying, at step 1425, an inverse transform to the enhanced trasnform coeeficients. At step 1430, the encoding device generates an image based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.

[0201] FIG. 15 illustrates an embodiment of a quantization step and inverse transform enhancement applied in a coding method 1500 implemented by a decoding device. At step 1505, thedecoding device receives transform coefficients. At step 1510, the decoding device generates a reconstructed transform. The decoding device generates the reconstructed transform by: applying, at step 1515, a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP; applying, at step 1520, an enhancement to the transform coeeficients; and applying, at step 1525, an inverse transform to the enhanced trasnform coeeficients. At step 1530, the decoding device generates an image based on the reconstructed transform. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.

[0202] FIG. 16 illustrates an embodiment of layers 1600 applied in a neural network-based loop filter, where CReLU components replace ReLU or PReLU components as the activation function. In particular, the layers 1400 include a first layer 1005, a second layer 1010, and athird layer 1015a. In the third layer 1015a, CReLU is applied as the activation function instead of ReLU components or the PReLU components of FIG. 10.

[0203] In this regard, during the utilization of neural network-based coding tools, it has been noticed that within the first few convolution layers, there is a phenomenon where each filter is nearly mirrored by another filter, albeit in the opposite phase. This observation suggests that early layers of deep CNNs adeptly capture both positive and negative phase information by learning pairs or groups of filters with negative correlations. Consequently, this implies redundancy within the filters of the first few convolutional layers.

[0204] To address this redundancy, an activation scheme known as CReLU is proposed to remove this redundancy for the first few convolution layers. CReLU effectively preserves both positive and negative phase information while ensuring non-saturated non-linearity. As such, the distinct characteristic of CReLU enables a mathematical depiction of convolution layers in terms of a reconstruction property, as represented by: CReLU(x) = concat( ReLU(x), ReLU(-x) ), where x is the output tensor of previous layer. Implementing CReLU results in doubling the output channel size for CReLU operation (the input channel size for the subsequent layer).

[0205] In an embodiment, a CReLU activation operation is applied to one or more layers. As a result, the CReLU activation operation doubles the output channel size without increasing the number of parameters for a previous layer. In another embodiment, to keep the output channel size of CReLU operation unchanged, the output channel size of previous layer is cut to half, results in decreasing the number of parameters for some or all previous layers.

[0206] In an embodiment, CReLU may also be represented as: CReLU(x) = ReLU( concat( x, - x) ), where x is the output tensor of previous layer. In this case, x and -x are concatenated and the concatenated tensor is used as the input to the ReLU function. If it is desired to keep the output channel size of ReLU operation unchanged, the output channel size of previous layer is cut to half, resulting in decreasing the number of parameters for some or all previous layers. As a result, redundancy within the filters of convolutional layers is removed, channel size is doubled, and model performance is improved.

[0207] FIG. 17 illustrates an embodiment of a coding method 1700 implemented by an encoding device. At step 1705, the encoding device receives transform coefficients. At step 1710, the encoding device applies a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform. At step 1715, the encoding device generates a bitstream based on the reconstructed transform. At step 1720, the encoding device applies a CReLU activation function during the quantization process. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.

[0208] FIG. 18 illustrates an embodiment of coding method 1800 implemented by an decoding device. At step 1805, the decoding device receives transform coefficients. At step 1810, the decoding device applies a QS to the transform coefficients during a quantization process instead of applying a base QP or a slice QP to obtain a reconstructed transform. At step 1815, the decoding device generates an image based on the reconstructed transform. At step 1820, the encoding device applies a CReLU activation function during the quantization process. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-readable medium.(3) Low-Rank Decomposition

[0209] Disclosed herein are techniques that improve reconstruction accuracy of inverse transforms by applying a low-rank decomposition to a convolutional layer. In particular, the method decomposes a convolutional layer to produce a convolutional layer having a fewer number of filters. By applying the low-rank decomposition, overly complex processing may be avoided during reconstruction of the inverse transform. Thus, usage of the processor, memory, and / or network resources may be reduced at both the encoder and the decoder. Thus, the coder / decoder (a.k.a., “codec”) in video coding is improved relative to current codecs. As a practical matter, the improvedvideo coding process offers the user a better user experience when videos are sent, received, and / or viewed.

[0210] FIG. 19 illustrates an embodiment of a low-rank decomposition applied in a coding method 1900 and implemented by an encoding device in a CNN. At step 1905, a reconstructed block is input into a convolution layer. At step 1910, a low-rank decomposition is applied to the reconstructed block within the convolution layer. At step 1915, a fdtered reconstructed block is generated based on application of the low-rank decomposition to the reconstructed block within the convolution layer. At step 1920, a bitstream is generated based on the filtered reconstructed block. As a result of the low-rank decomposition, a matrix input to the convolution layer is flattened such that the number of coefficients in the matrix is decreased, and the number of computations performed in the convolution layer is decreased.

[0211] In the embodiment, a low-rank decomposition is applied to a convolution layer. For example, for a convolutional layer having dimensions (m, n, k, k), where m and n are the number of input and output channels and k is the size of a convolution filter, the 4D tensor W is flattened to a 2D matrix of shape (mk2, n). The rank of the flattened matrix is no more than min(mk2, n). This 2D matrix can be decomposed to U E / ?mk2 xran(j yTG Rr nwhere r represents the rank of the 2D matrix. The decomposed U, VTmatrices can be reshaped back to 4D tensor U G Rmxrxkxkand vTG Rrxnor VTE Rr'nxlxl.

[0212] Decomposing a convolutional layer produces a convolutional layer having a fewer number of filters. For example, decomposing a convolutional layer produces a convolutional layer U with r convolution filters and a linear projection layer VT, which can also be seen as a 1 x 1 convolutional layer V1with r convolution filters. After low-rank decomposition, the complexity (MAC count) is changed from nxmxk2xwxh to (n+mxk2) r*wxh where w and h are the width and height of input tensor. The complexity is reduced if r < (nxmxk2) / (n+mxk2). This breakdown of the matrix to fewer components results in more efficient encoder and decoder operations.

[0213] FIG. 20 illustrates an embodiment of a low-rank decomposition applied in a coding method and implemented by a decoding device in a CNN. At step 2005, a reconstructed block is input into a convolution layer. At step 2010, a low-rank decomposition is applied to the reconstructed block within the convolution layer. At step 2015, a filtered reconstructed block is generated based on application of the low-rank decomposition to the reconstructed block within the convolution layer. At step 2020, an image is generated based on the filtered reconstructed block.

[0214] In another embodiment, a low-rank decomposition may be applied to a fully-connected (FC) neural network or a multi-layer perceptron (MLP) layer. In this regard, an FC neural network can be represented as f(x) = xW, where W represents weight matrices with dimension (n, w), and where x is the input tensor with dimension (w, h). The weight matrix W can be decomposed as UVTwhere U G Rnxrand VTG Rrxw

[0215] A similar approach may be applied to MLP mixer layers, where each weight matrix can be decomposed in the same manner. Decomposing an FC layer produces a convolutional layer having a fewer number of filters. For example, decomposing an FC layer produces an FC layer U and a linear projection layer VT. After the low-rank decomposition, the complexity (MAC count) can be changed from nxwxh to (n+w)xrxh. The complexity is reduced if r < (n*w) / (n+w).

[0216] In another embodiment, a low-rank decomposition may be applied to a multi-head attention (MHA) layer. A p-head attention layer learns p attention mechanisms on a key, value, and query (K, V, Q) of each input token: MHA(Q, K, V) = Concat(headi, . . . , headp)W°.

[0217] Each head is computed by: head; = Attention( QWQ KW^l) = softmax( ) VW^, where d isthe hidden dimension.

[0218] The weightsi 6 (1, 2, . . ., p) can be decomposed to UVTin the same manner.

[0219] To further decrease the model complexity, a low-rank decomposition operation is provided to convert the weight of some or all layers to its U, VTmatrices form and change the layer definition accordingly. The parameter r can be determined arbitrarily, or it can be determined based on a predefined or adaptive RD measurement.

[0220] FIG. 21 is a schematic diagram of an apparatus 2100 (e.g., a network apparatus, a network node, a network router, a router, etc.). The apparatus 2100 is suitable for implementing the disclosed embodiments as described herein. The apparatus 2100 comprises ingress ports / ingress means 2110 (a.k.a., upstream ports) and receiver units (Rx) / receiving means 2120 for receiving data; a processor, logic unit, or central processing unit (CPU) / processing means 2130 to process the data; transmitter units (Tx) / transmitting means 2140 and egress ports / egress means 2150 (a.k.a., downstream ports) for transmitting the data; and a memory / memory means 2160 for storing the data. The apparatus 2100 may also comprise optical-to-electrical (OE) components and electrical-to-optical (EO) components coupled to the ingress ports / ingress means 2110, the receiver units / receiving means 2120, the transmitter units / transmitting means 2140, and the egress ports / egress means 2150 for egress or ingress of optical or electrical signals.

[0221] The processor / processing means 2130 is implemented by hardware and software. The processor / processing means 2130 may be implemented as one or more CPU chips, cores (e.g., as a multi-core processor), field-programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), and digital signal processors (DSPs). The processor / processing means 2130 is in communication with the ingress ports / ingress means 2110, receiver units / receiving means 2120, transmitter units / transmitting means 2140, egress ports / egress means 2150, and memory / memory means 2160. The processor / processing means 2130 comprises a module 2170. The module 2170 is able to implement the methods disclosed herein. The inclusion of the module 2170 therefore provides a substantial improvement to the functionality of the apparatus 2100 and effects a transformation of the apparatus 2100 to a different state. Alternatively, the module 2170 is implemented as instructions stored in the memory / memory means 2160 and executed by the processor / processing means 2130.

[0222] The apparatus 2100 may also include input and / or output (VO) devices or I / O means 2115 for communicating data to and from a user. The VO devices or I / O means 2115 may include output devices such as a display for displaying video data, speakers for outputting audio data, etc. The VO devices or VO means 2115 may also include input devices, such as a keyboard, mouse, trackball, etc., and / or corresponding interfaces for interacting with such output devices.

[0223] The memory / memory means 2160 comprises one or more disks, tape drives, and solid- state drives and may be used as an over-flow data storage device, to store programs when such programs are selected for execution, and to store instructions and data that are read during program execution. The memory / memory means 2160 may be volatile and / or non-volatile and may be readonly memory (ROM), random access memory (RAM), ternary content-addressable memory (TCAM), and / or static random-access memory (SRAM).

[0224] It should also be understood that the steps of the exemplary methods set forth herein are not necessarily required to be performed in the order described, and the order of the steps of such methods should be understood to be merely exemplary. Likewise, additional steps may be included in such methods, and certain steps may be omitted or combined, in methods consistent with various embodiments of the present disclosure.

[0225] While several embodiments have been provided in the present disclosure, it should be understood that the disclosed systems and methods might be embodied in many other specific forms without departing from the spirit or scope of the present disclosure. The present examples are to be considered as illustrative and not restrictive, and the intention is not to be limited to the details given herein. For example, the various elements or components may be combined or integrated in another system or certain features may be omitted, or not implemented.

[0226] In addition, techniques, systems, subsystems, and methods described and illustrated in the various embodiments as discrete or separate may be combined or integrated with other systems, modules, techniques, or methods without departing from the scope of the present disclosure. Other items shown or discussed as coupled or directly coupled or communicating with each other may be indirectly coupled or communicating through some interface, device, or intermediate component whether electrically, mechanically, or otherwise. Other examples of changes, substitutions, and alterations are ascertainable by one skilled in the art and could be made without departing from the spirit and scope disclosed herein.

Claims

CLAIMSWhat is claimed is:

1. A coding method implemented by a decoding device in a convolutional neural network (CNN), the method comprising: receiving transform coefficients; generating a reconstructed transform by: applying a quantization step (QS) to the transform coefficients during a quantization process instead of applying a base quantization parameter (QP) or a slice QP; applying an enhancement to the transform coefficients; and applying an inverse transform to the enhanced transform coefficients; and generating an image based on the reconstructed transform.

2. The coding method of claim 1, wherein applying the QS comprises applying an activation function during the quantization process.

3. The coding method of any of claims 1-2, wherein the activation function comprises a concatenated rectified linear unit (CReLU).

4. The coding method of any of claims 1-3, further comprising applying the CReLU instead of a rectified linear unit (ReLU) or a parametric ReLU (PReLU).

5. The coding method of any of claims 1-4, wherein applying the enhancement comprises applying a plurality of enhancement processes, and wherein applying the inverse transform comprises applying a plurality of inverse transform operations.

6. The coding method of any of claims 1-5, further comprising performing an enhancement process before the inverse transform is applied.

7. The coding method of any of claims 1-6, wherein the plurality of enhancement processes comprises a first coefficient enhancement, a second coefficient enhancement, avertical enhancement, and a horizontal enhancement, and wherein the plurality of inverse transform operations comprises an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

8. The coding method of any of claims 1-7, further comprising: performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

9. A coding method implemented by an encoding device in a convolutional neural network (CNN), the method comprising: receiving transform coefficients; generating a reconstructed transform by: applying a quantization step (QS) to the transform coefficients during a quantization process instead of applying a base quantization parameter (QP) or a slice QP; applying an enhancement to the transform coefficients; and applying an inverse transform to the enhanced transform coefficients; and generating a bitstream based on the reconstructed transform.

10. The coding method of claim 9, wherein applying the QS comprises applying an activation function during the quantization process.

11. The coding method of any of claims 9-10, wherein the activation function comprises a concatenated rectified linear unit (CReLU).

12. The coding method of any of claims 9-11, further comprising applying the CReLU instead of a rectified linear unit (ReLU) or a parametric ReLU (PReLU).

13. The coding method of any of claims 9-12, wherein applying the enhancement comprises applying a plurality of enhancement processes, and wherein applying the inverse transform comprises applying a plurality of inverse transform operations.

14. The coding method of any of claims 9-13, further comprising performing an enhancement process before the inverse transform is applied.

15. The coding method of any of claims 9-14, wherein the plurality of enhancement processes comprises a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and wherein the plurality of inverse transform operations comprises an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

16. The coding method of any of claims 9-15, further comprising: performing the first coefficient enhancement before performing the inverse secondary transform; performing the second coefficient enhancement after performing the inverse secondary transform; performing the vertical enhancement after performing the inverse primary vertical transform; and performing the horizontal enhancement after performing the inverse primary horizontal transform.

17. A method of coding implemented by a decoding device in a convolutional neural network (CNN), the method comprising:inputting a reconstructed block into a convolution layer; applying a low-rank decomposition to the reconstructed block within the convolution layer; generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and generating an image based on the filtered reconstructed block.

18. The method of claim 17, wherein the reconstructed block comprises a first matrix that is output from a previous layer, wherein the method further comprises applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and wherein the second matrix has a lower weight than the first matrix.

19. The method of any of claims 17-18, wherein the first matrix comprises a first number of coefficients and the second matrix comprises a second number of coefficients, and wherein the second number is less than the first number.

20. A method of coding implemented by an encoding device in a convolutional neural network (CNN), the method comprising: inputting a reconstructed block into a convolution layer; applying a low-rank decomposition to the reconstructed block within the convolution layer; generating a filtered reconstructed block based on application of the low-rank decomposition to the reconstructed block within the convolution layer; and generating a bitstream based on the filtered reconstructed block.

21. The method of claim 20, wherein the reconstructed block comprises a first matrix that is output from a previous layer, wherein the method further comprises applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layerto obtain a second matrix, and wherein the second matrix has a lower weight than the first matrix.

22. The method of any of claims 20-21, wherein the first matrix comprises a first number of coefficients and the second matrix comprises a second number of coefficients, and wherein the second number is less than the first number.

23. A coding device in a convolutional neural network (CNN), comprising: a memory storing instructions; and one or more processors coupled to the memory, the one or more processors configured to execute the instructions to cause the coding device to perform the method in any of claims 1-22.

24. A non-transitory computer readable medium comprising a computer program product for use by a computing device, the computer program product comprising computer executable instructions stored on the non-transitory computer readable medium that, when executed by one or more processors, cause the computing device to execute the method in any of claims 1-22.

25. A coding device comprising: means for receiving transform coefficients; means for generating a reconstructed transform comprising: means for applying a quantization step (QS) to the transform coefficients during a quantization process instead of applying a base quantization parameter (QP) or a slice QP; means for applying an enhancement to the transform coefficients; and means for applying an inverse transform to the enhanced transform coefficients; and means for generating an image based on the reconstructed transform.

26. The computing device of claim 25, further comprising means for applying an activation function during the quantization process.

27. The computing device of any of claims 25-26, wherein the activation function comprises a concatenated rectified linear unit (CReLU).

28. The computing device of any of claims 25-27, further comprising means for applying the CReLU instead of a rectified linear unit (ReLU) or a parametric ReLU (PReLU).

29. The computing device of any of claims 25-28, further comprising: means for applying a plurality of enhancement processes; and means for applying a plurality of inverse transform operations.

30. The computing device of any of claims 25-29, further comprising means for performing an enhancement process before the inverse transform is applied.

31. The computing device of any of claims 25-30, wherein the plurality of enhancement processes comprises a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and wherein the plurality of inverse transform operations comprises an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

32. The computing device of claim 25-31, further comprising: means for performing the first coefficient enhancement before performing the inverse secondary transform; means for performing the second coefficient enhancement after performing the inverse secondary transform; means for performing the vertical enhancement after performing the inverse primary vertical transform; and means for performing the horizontal enhancement after performing the inverse primary horizontal transform.

33. A coding device comprising: means for receiving transform coefficients; means for generating a reconstructed transform comprising: means for applying a quantization step (QS) to the transform coefficients during a quantization process instead of applying a base quantization parameter (QP) or a slice QP; means for applying an enhancement to the transform coefficients; and means for applying an inverse transform to the enhanced transform coefficients; and means for generating a bit stream based on the reconstructed transform.

34. The computing device of claim 33, wherein means for applying the QS comprises means for applying an activation function during the quantization process.

35. The computing device of any of claims 33-34, wherein the activation function comprises a concatenated rectified linear unit (CReLU).

36. The computing device of any of claims 33-35, further comprising means for applying the CReLU instead of a rectified linear unit (ReLU) or a parametric ReLU (PReLU).

37. The computing device of any of claims 33-36, further comprising: means for applying a plurality of enhancement processes; and means for applying a plurality of inverse transform operations.

38. The computing device of any of claims 33-37, further comprising means for performing an enhancement process before the inverse transform is applied.

39. The computing device of any of claims 33-38, wherein the plurality of enhancement processes comprises a first coefficient enhancement, a second coefficient enhancement, a vertical enhancement, and a horizontal enhancement, and wherein the plurality of inversetransform operations comprises an inverse secondary transform, an inverse primary vertical transform, and an inverse primary horizontal transform.

40. The computing device of any of claims 33-39, further comprising: means for performing the first coefficient enhancement before performing the inverse secondary transform; means for performing the second coefficient enhancement after performing the inverse secondary transform; means for performing the vertical enhancement after performing the inverse primary vertical transform; and means for performing the horizontal enhancement after performing the inverse primary horizontal transform.

41. A computing device, comprising: means for inputting a reconstructed block into a convolution layer; means for applying a low-rank decomposition to the reconstructed block within the convolution layer; means for generating a filtered reconstructed block based on application of the low- rank decomposition to the reconstructed block within the convolution layer; and means for generating an image based on the filtered reconstructed block.

42. The computing device of claim 41, wherein the reconstructed block comprises a first matrix that is output from a previous layer, wherein the computing device further comprises means for applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and wherein the second matrix has a lower weight than the first matrix.

43. The computing device of any of claims 41-42, wherein the first matrix comprises a first number of coefficients and the second matrix comprises a second number of coefficients, and wherein the second number is less than the first number.

44. A computing device, comprising: means for inputting a reconstructed block into a convolution layer; means for applying a low-rank decomposition to the reconstructed block within the convolution layer; means for generating a filtered reconstructed block based on application of the low- rank decomposition to the reconstructed block within the convolution layer; and means for generating a bitstream based on the filtered reconstructed block.

45. The computing device of claim 44, wherein the reconstructed block comprises a first matrix that is output from a previous layer, wherein the computing device further comprises means for applying the low-rank decomposition by multiplying the first matrix by a weight matrix within the convolution layer to obtain a second matrix, and wherein the second matrix has a lower weight than the first matrix.

46. The computing device of any of claims 44-45, wherein the first matrix comprises a first number of coefficients and the second matrix comprises a second number of coefficients, and wherein the second number is less than the first number.

Images