Method, apparatus, and medium for visual data processing
By processing visual data with a neural network-based model that considers both local and global features, the method addresses inefficiencies in existing visual data compression, achieving improved coding efficiency.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- DOUYIN VISION CO LTD
- Filing Date
- 2026-01-23
- Publication Date
- 2026-07-30
AI Technical Summary
Existing neural network-based visual data compression technologies suffer from inefficiencies in coding due to redundancy in feature representation, limiting their coding efficiency.
A method and apparatus that utilize a neural network-based model to process visual data by considering both local and global features in the feature representation, reducing redundancy and improving coding efficiency.
The proposed method effectively reduces redundancy in feature representation, enhancing the coding efficiency of visual data processing.
Smart Images

Figure CN2026074732_30072026_PF_FP_ABST
Abstract
Description
METHOD, APPARATUS, AND MEDIUM FOR VISUAL DATA PROCESSINGFIELDS
[0001] Cases of the present disclosure relates generally to visual data processing techniques, and more particularly, to visual data processing with a neural network (NN) -based model.BACKGROUND
[0002] The past decade has witnessed the rapid development of deep learning in a variety of areas, especially in computer vision and image processing. Neural network was invented originally with the interdisciplinary research of neuroscience and mathematics. It has shown strong capabilities in the context of non-linear transform and classification. Neural network-based image / video compression technology has made significant progress during the past half decade. It is reported that the latest neural network-based image compression algorithm achieves comparable rate-distortion (R-D) performance with Versatile Video Coding (VVC) . With the performance of neural image compression continually improving, neural network-based visual data compression has become an actively developing research area. However, coding efficiency of neural network-based visual data processing is generally expected to be further improved.SUMMARY
[0003] Cases of the present disclosure provide a solution for visual data processing.
[0004] In a first aspect, a method for visual data processing is proposed. The method comprises: performing a conversion between visual data and a bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.
[0005] Compared with the conventional solution, the proposed method can advantageously jointly consider the local feature and the global feature associated with the feature representation that is associated with the visual data so as to reduce the redundancy in the feature representation space, and thus improving the coding efficiency.
[0006] In a second aspect, an apparatus for visual data processing is proposed. The apparatus comprises a processor and a non-transitory memory with instructions thereon. The instructions upon execution by the processor, cause the processor to perform a method in accordance with the first aspect of the present disclosure.
[0007] In a third aspect, a non-transitory computer-readable storage medium is proposed. The non-transitory computer-readable storage medium stores instructions that cause a processor to perform a method in accordance with the first aspect of the present disclosure.
[0008] In a fourth aspect, another non-transitory computer-readable recording medium is proposed. The non-transitory computer-readable recording medium stores a bitstream of a visual data which is generated by a method performed by an apparatus for visual data processing. The method comprises: generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.
[0009] In a fifth aspect, a method for storing a bitstream of a visual data is proposed. The method comprises: generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data; and storing the bitstream in a non-transitory computer-readable recording medium.
[0010] In a sixth aspect, a computer program is proposed. The computer program is for performing a method in accordance with the first aspect of the present disclosure when the computer program runs on a computer.
[0011] In a seventh aspect, a computer program product is proposed. The computer program product is tangibly stored in a computer storage medium and comprises computer-executable instructions that, when executed by a device, cause the device to perform a method in accordance with the first aspect of the present disclosure.
[0012] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Through the following detailed description with reference to the accompanying drawings, the above and other objectives, features, and advantages of example cases of the present disclosure will become more apparent. In the example cases of the present disclosure, the same reference numerals usually refer to the same components.
[0014] Fig. 1A illustrates a block diagram of an example visual data coding system in accordance with some cases of the present disclosure;
[0015] Fig. 1B illustrates a typical transform coding scheme;
[0016] Fig. 2 illustrates an image from the Kodak dataset and different representations of the image;
[0017] Fig. 3 illustrates a network architecture of an autoencoder implementing the hyperprior model;
[0018] Fig. 4 illustrates a block diagram of a combined model;
[0019] Fig. 5 illustrates an encoding process;
[0020] Fig. 6 illustrates a decoding process;
[0021] Fig. 7A illustrates a structure of the multistage context model;
[0022] Fig. 7B illustrates a shape of the context model;
[0023] Fig. 8A illustrates an example structure of a frequency decomposition transformer (FDT) module in accordance with some cases of the present disclosure;
[0024] Fig. 8B illustrates a frequency (shifted) window-based self-attention (F (S) WSA) module of the FDT module in accordance with some cases of the present disclosure;
[0025] Fig. 8C illustrates a frequency domain feed-forward network (FDFFN) of the FDT module in accordance with some cases of the present disclosure;
[0026] Fig. 8D illustrates a low frequency / high frequency- (shifted) window-based self-attention (LF / HF-(S)WSA) of the FDT module in accordance with some cases of the present disclosure;
[0027] Fig. 9 illustrates an overall framework for visual data coding in accordance with some cases of the present disclosure;
[0028] Fig. 10 illustrates a flowchart of a method for visual data processing in accordance with some cases of the present disclosure;
[0029] Fig. 11 illustrates an overview of a neural video codec based on a Dual-Scale Transformer (DST) -based contextual encoder-decoder module in accordance with some cases of the present disclosure;
[0030] Fig. 12A illustrates an overview of the DST block in accordance with some cases of the present disclosure;
[0031] Fig. 12B illustrates a Global-Local (Shifted) Window-based Self-Attention (GL (S) WSA module of the DST block in accordance with some cases of the present disclosure;
[0032] Fig. 12C illustrates a Cross-Gated Feed-Forward Network (CGFFN) of the DST block in accordance with some cases of the present disclosure;
[0033] Fig. 12D illustrates a Global / Local (Shifted) Window based Self-Attention (Global / Local (S) WSA) of the DST block in accordance with some cases of the present disclosure;
[0034] Fig. 13A illustrates an asynchronous training strategy on multiple GPU;
[0035] Fig. 13B illustrates a synchronous training strategy on multiple GPU in accordance with some cases of the present disclosure; and
[0036] Fig. 14 illustrates a block diagram of a computing device in which various cases of the present disclosure can be implemented.
[0037] Throughout the drawings, the same or similar reference numerals usually refer to the same or similar elements.DETAILED DESCRIPTION
[0038] Principle of the present disclosure will now be described with reference to some cases. It is to be understood that these cases are described only for the purpose of illustration and help those skilled in the art to understand and implement the present disclosure, without suggesting any limitation as to the scope of the disclosure. The disclosure described herein can be implemented in various manners other than the ones described below.
[0039] In the following description and claims, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure belongs.
[0040] References in the present disclosure to “one case, ” “an case, ” “an example case, ” and the like indicate that the case described may include a particular feature, structure, or characteristic, but it is not necessary that every case includes the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same case. Further, when a particular feature, structure, or characteristic is described in connection with an example cases, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other cases whether or not explicitly described.
[0041] It shall be understood that although the terms “first” and “second” etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and similarly, a second element could be termed a first element, without departing from the scope of example cases. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.
[0042] The terminology used herein is for the purpose of describing particular cases only and is not intended to be limiting of example cases. As used herein, the singular forms “a” , “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” , “comprising” , “has” , “having” , “includes” and / or “including” , when used herein, specify the presence of stated features, elements, and / or components etc., but do not preclude the presence or addition of one or more other features, elements, components and / or combinations thereof. Example Environment
[0043] Fig. 1A is a block diagram that illustrates an example visual data coding system 100 that may utilize the techniques of this disclosure. As shown, the visual data coding system 100 may include a source device 110 and a destination device 120. The source device 110 can be also referred to as a visual data encoding device, and the destination device 120 can be also referred to as a visual data decoding device. In operation, the source device 110 can be configured to generate encoded visual data and the destination device 120 can be configured to decode the encoded visual data generated by the source device 110. The source device 110 may include a visual data source 112, a visual data encoder 114, and an input / output (I / O) interface 116.
[0044] The visual data source 112 may include a source such as a visual data capture device. Examples of the visual data capture device include, but are not limited to, an interface to receive visual data from a visual data content provider, a computer graphics system for generating visual data, and / or a combination thereof.
[0045] The visual data may comprise one or more pictures. The visual data encoder 114 encodes the visual data from the visual data source 112 to generate a bitstream. The bitstream may include a sequence of bits that form a coded representation of the visual data. The bitstream may include coded pictures and associated data. The coded picture is a coded representation of a picture. The associated data may include sequence parameter sets, picture parameter sets, and other syntax structures. The I / O interface 116 may include a modulator / demodulator and / or a transmitter. The encoded visual data may be transmitted directly to destination device 120 via the I / O interface 116 through the network 130A. The encoded visual data may also be stored onto a storage medium / server 130B for access by destination device 120.
[0046] The destination device 120 may include an I / O interface 126, a visual data decoder 124, and a display device 122. The I / O interface 126 may include a receiver and / or a modem. The I / O interface 126 may acquire encoded visual data from the source device 110 or the storage medium / server 130B. The visual data decoder 124 may decode the encoded visual data. The display device 122 may display the decoded visual data to a user. The display device 122 may be integrated with the destination device 120, or may be external to the destination device 120 which is configured to interface with an external display device.
[0047] The visual data encoder 114 and the visual data decoder 124 may operate according to a visual data compression standard, such as the High Efficiency Video Coding (HEVC) standard, Versatile Video Coding (VVC) standard and other current and / or further standards. 1. Brief Summary A neural network-based video compression method for better compression performance with the frequency decomposition transformer. In the present disclosure, frequency decomposition transformer (FDT) block is proposed to capture low-frequency and high-frequency components to generate more compact latent representations. 2. Introduction The past decade has witnessed the rapid development of deep learning in a variety of areas, especially in computer vision and image processing. Inspired from the great success of deep learning technology to computer vision areas, many researchers have shifted their attention from conventional image / video compression techniques to neural image / video compression technologies. Neural network was invented originally with the interdisciplinary research of neuroscience and mathematics. It has shown strong capabilities in the context of non-linear transform and classification. Neural network-based image / video compression technology has made significant progress during the past half decade. It is reported that the latest neural network-based image compression algorithm achieves comparable R-D performance with Versatile Video Coding (VVC) , the latest video coding standard developed by Joint Video Experts Team (JVET) with experts from MPEG and VCEG. With the performance of neural image compression continually improving, neural network-based video compression has become an actively developing research area. However, neural network-based video coding still remains in its infancy due to the inherent difficulty of the problem. 2.1. Image / Video compression Image / video compression usually refers to the computing technology that compresses image / video into binary code to facilitate storage and transmission. The binary codes may or may not support lossless reconstructing the original image / video, termed lossless compression and lossy compression. Most of the efforts are devoted to lossy compression since lossless reconstruction is not necessary in most scenarios. Usually the performance of image / video compression algorithms is evaluated from two aspects, i.e. compression ratio and reconstruction quality. Compression ratio is directly related to the number of binary codes, the less the better; Reconstruction quality is measured by comparing the reconstructed image / video with the original image / video, the higher the better. Image / video compression techniques can be divided into two branches, the classical video coding methods and the neural-network-based video compression methods. Classical video coding schemes adopt transform-based solutions, in which researchers have exploited statistical dependency in the latent variables (e.g., DCT or wavelet coefficients) by carefully hand-engineering entropy codes modeling the dependencies in the quantized regime. Neural network-based video compression is in two flavors, neural network-based coding tools and end-to-end neural network-based video compression. The former is embedded into existing classical video codecs as coding tools and only serves as part of the framework, while the latter is a separate framework developed based on neural networks without depending on classical video codecs. In the last three decades, a series of classical video coding standards have been developed to accommodate the increasing visual content. The international standardization organizations ISO / IEC has two expert groups namely Joint Photographic Experts Group (JPEG) and Moving Picture Experts Group (MPEG) , and ITU-T also has its own Video Coding Experts Group (VCEG) which is for standardization of image / video coding technology. The influential video coding standards published by these organizations include JPEG, JPEG 2000, H. 262, H.264 / AVC and H. 265 / HEVC. After H. 265 / HEVC, the Joint Video Experts Team (JVET) formed by MPEG and VCEG has been working on a new video coding standard Versatile Video Coding (VVC) . The first version of VVC was released in July 2020. An average of 50%bitrate reduction is reported by VVC under the same visual quality compared with HEVC. Neural network-based image / video compression is not a new solution since there were a number of researchers working on neural network-based image coding. But the network architectures were relatively shallow, and the performance was not satisfactory. Benefit from the abundance of data and the support of powerful computing resources, neural network-based methods are better exploited in a variety of applications. At present, neural network-based image / video compression has shown promising improvements, confirmed its feasibility. Nevertheless, this technology is still far from mature and a lot of challenges need to be addressed. 2.2. Neural networks Neural networks, also known as artificial neural networks (ANN) , are the computational models used in machine learning technology which are usually composed of multiple processing layers and each layer is composed of multiple simple but non-linear basic computational units. One benefit of such deep networks is believed to be the capacity for processing data with multiple levels of abstraction and converting data into different kinds of representations. Note that these representations are not manually designed; instead, the deep network including the processing layers is learned from massive data using a general machine learning procedure. Deep learning eliminates the necessity of handcrafted representations, and thus is regarded useful especially for processing natively unstructured data, such as acoustic and visual signal, whilst processing such data has been a longstanding difficulty in the artificial intelligence field. 2.3. Neural networks for image compression Existing neural networks for image compression methods can be classified in two categories, i.e., pixel probability modeling and auto-encoder. The former one belongs to the predictive coding strategy, while the latter one is the transform-based solution. Sometimes, these two methods are combined together in literature. 2.3.1. Pixel Probability Modeling According to Shannon’s information theory, the optimal method for lossless coding can reach the minimal coding rate -log2p (x) where p (x) is the probability of symbol x. A number of lossless coding methods were developed in literature and among them arithmetic coding is believed to be among the optimal ones. Given a probability distribution p (x) , arithmetic coding ensures that the coding rate to be as close as possible to its theoretical limit -log2p (x) without considering the rounding error. Therefore, the remaining problem is to how to determine the probability, which is however very challenging for natural image / video due to the curse of dimensionality. Following the predictive coding strategy, one way to model p (x) is to predict pixel probabilities one by one in a raster scan order based on previous observations, where x is an image. p(x) =p (x1) p (x2|x1) …p (xi|x1, …, xi-1) …p (xm×n|x1, …, xm×n-1) (2-1) where m and n are the height and width of the image, respectively. The previous observation is also known as the context of the current pixel. When the image is large, it can be difficult to estimate the conditional probability, thereby a simplified method is to limit the range of its context. p(x) =p (x1) p (x2|x1) …p (xi|xi-k, …, xi-1) …p (xm×n|xm×n-k, …, xm×n-1) (2-2) where k is a pre-defined constant controlling the range of the context. It should be noted that the condition may also take the sample values of other color components into consideration. For example, when coding the RGB color component, R sample is dependent on previously coded pixels (including R / G / B samples) , the current G sample may be coded according to previously coded pixels and the current R sample, while for coding the current B sample, the previously coded pixels and the current R and G samples may also be taken into consideration. Neural networks were originally introduced for computer vision tasks and have been proven to be effective in regression and classification problems. Therefore, it has been proposed using neural networks to estimate the probability of p (xi) given its context x1, x2, …, xi-1. The pixel probability is proposed for binary images, i.e., xi∈{-1, +1} . The neural autoregressive distribution estimator (NADE) is designed for pixel probability modeling, where there is a feed-forward network with a single hidden layer. In a similar work, the feed-forward network also has connections skipping the hidden layer, and the parameters are also shared. Experiments have been performed on the binarized MNIST dataset. NADE is extended to a real-valued model RNADE, where the probability p (xi|x1, …, xi-1) is derived with a mixture of Gaussians. Their feed-forward network also has a single hidden layer, but the hidden layer is with rescaling to avoid saturation and uses rectified linear unit (ReLU) instead of sigmoid. NADE and RNADE are improved by reorganizing the order of the pixels and with deeper neural networks. Designing advanced neural networks plays an important role in improving pixel probability modeling. Multi-dimensional long short-term memory (LSTM) is proposed, which is working together with mixtures of conditional Gaussian scale mixtures for probability modeling. LSTM is a special kind of recurrent neural networks (RNNs) and has proven to be good at modeling sequential data. The spatial variant of LSTM is used for images later. Several different neural networks are studied, including RNNs and CNNs namely PixelRNN and PixelCNN, respectively. In PixelRNN, two variants of LSTM, called row LSTM and diagonal BiLSTM are proposed, where the latter is specifically designed for images. PixelRNN incorporates residual connections to help train deep neural networks with up to 12 layers. In PixelCNN, masked convolutions are used to suit for the shape of the context. Comparing with previous works, PixelRNN and PixelCNN are more dedicated to natural images: pixels are considered as discrete values (e.g., 0, 1, …, 255) and predict a multinomial distribution over the discrete values; color images are deal with in RGB color space; good performance is shown on large-scale image dataset ImageNet. Gated PixelCNN is proposed to improve the PixelCNN and achieves comparable performance with PixelRNN but with much less complexity. PixelCNN++ is proposed with the following improvements upon PixelCNN: a discretized logistic mixture likelihood is used rather than a 256-way multinomial distribution; down-sampling is used to capture structures at multiple resolutions; additional short-cut connections are introduced to speed up training; dropout is adopted for regularization; RGB is combined for one pixel. PixelSNAIL is proposed, in which casual convolutions are combined with self-attention. Most of the above methods directly model the probability distribution in the pixel domain. Some researchers also attempt to model the probability distribution as a conditional one upon explicit or latent representations. That being said, it may be estimated where h is the additional condition and p (x) =p (h) p (x|h) , meaning the modeling is split into an unconditional one and a conditional one. The additional condition can be image label information or high-level representations. 2.3.2. Auto-encoder Auto-encoder originates from the well-known work proposed by Hinton and Salakhutdinov. The method is trained for dimensionality reduction and consists of two parts: encoding and decoding. The encoding part converts the high-dimension input signal to low-dimension representations, typically with reduced spatial size but a greater number of channels. The decoding part attempts to recover the high-dimension input from the low-dimension representation. Auto-encoder enables automated learning of representations and eliminates the need of hand-crafted features, which is also believed to be one of the most important advantages of neural networks. Fig. 1 illustrates a typical transform coding scheme. The original image x is transformed by the analysis network ga to achieve the latent representation y. The latent representation y is quantized and compressed into bits. The number of bits R is used to measure the coding rate. The quantized latent representation is then inversely transformed by a synthesis network gs to obtain the reconstructed image The distortion is calculated in a perceptual space by transforming x and with the function gp. It is intuitive to apply auto-encoder network to lossy image compression. It is only needed to encode the learned latent representation from the well-trained neural networks. However, it is not trivial to adapt auto-encoder to image compression since the original auto-encoder is not optimized for compression thereby not efficient by directly using a trained auto-encoder. In addition, there exist other major challenges: First, the low-dimension representation should be quantized before being encoded, but the quantization is not differentiable, which is required in backpropagation while training the neural networks. Second, the objective under compression scenario is different since both the distortion and the rate need to be taken into consideration. Estimating the rate is challenging. Third, a practical image coding scheme needs to support variable rate, scalability, encoding / decoding speed, interoperability. In response to these challenges, a number of researchers have been actively contributing to this area. The prototype auto-encoder for image compression is in Fig. 1B, which can be regarded as a transform coding strategy. The original image x is transformed with the analysis network y=ga (x) , where y is the latent representation which will be quantized and coded. The synthesis network will inversely transform the quantized latent representation back to obtain the reconstructed image The framework is trained with the rate-distortion loss function, i.e., where D is the distortion between x and R is the rate calculated or estimated from the quantized representation and λ is the Lagrange multiplier. It should be noted that D can be calculated in either pixel domain or perceptual domain. All existing research works follow this prototype and the difference might only be the network structure or loss function. In terms of network structure, RNNs and CNNs are the most widely used architectures. In the RNNs relevant category, a general framework was proposed for variable rate image compression using RNN. Binary quantization is used to generate codes and do not consider rate during training. The framework indeed provides a scalable coding functionality, where RNN with convolutional and deconvolution layers is reported to perform decently. Then an improved version was proposed by upgrading the encoder with a neural network similar to PixelRNN to compress the binary codes. The performance is reportedly better than JPEG on Kodak image dataset using MS-SSIM evaluation metric. The RNN-based solution was further improved by introducing hidden-state priming. In addition, an SSIM-weighted loss function is also designed, and spatially adaptive bitrates mechanism is enabled. Better results is achieved, than BPG on Kodak image dataset using MS-SSIM as evaluation metric. A general framework was designed for rate-distortion optimized image compression. Multiary quantization is used to generate integer codes and consider the rate during training, i.e. the loss is the joint rate-distortion cost, which can be MSE or others. Random uniform noise is added to stimulate the quantization during training and use the differential entropy of the noisy codes as a proxy for the rate. Generalized divisive normalization (GDN) is used as the network structure, which consists of a linear mapping followed by a nonlinear parametric normalization. The effectiveness of GDN on image coding is verified. An improved version was proposed, where 3 convolutional layers are used, each followed by a down-sampling layer and a GDN layer as the forward transform. Accordingly, 3 layers of inverse GDN are used, each followed by an up-sampling layer and convolution layer to stimulate the inverse transform. In addition, an arithmetic coding method is devised to compress the integer codes. The performance is reportedly better than JPEG and JPEG 2000 on Kodak dataset in terms of MSE. Furthermore, it was further improved by devising a scale hyper-prior into the auto-encoder. The latent representation y is transformed with a subnet ha to z=ha (y) and z will be quantized and transmitted as side information. Accordingly, the inverse transform is implemented with a subnet hs attempting to decode from the quantized side information to the standard deviation of the quantized which will be further used during the arithmetic coding of On the Kodak image set, their method is slightly worse than BPG in terms of PSNR. The structures were further explored in the residue space by introducing an autoregressive model to estimate both the standard deviation and the mean. In the latest work Gaussian mixture model was used to further remove redundancy in the residue. The reported performance is on par with VVC on the Kodak image set using PSNR as evaluation metric. 2.3.3. Hyper Prior Model In the transform coding approach to image compression, the encoder subnetwork (section 2.3.2) transforms the image vector x using a parametric analysis transform into a latent representation y, which is then quantized to form Because is discrete-valued, it can be lossless compressed using entropy coding techniques such as arithmetic coding and transmitted as a sequence of bits. Fig. 2 illustrates an image from the Kodak dataset and different representations of the image. As evident from the middle left and middle right image of Fig. 2, there are significant spatial dependencies among the elements of Notably, their scales (middle right image) appear to be coupled spatially. An additional set of random variables are introduced to capture the spatial dependencies and to further reduce the redundancies. In this case the image compression network is depicted in Fig. 3. In Fig. 3, the left hand of the models is the encoder ga and decoder gs (explained in section 2.3.2) . The right-hand side is the additional hyper encoder ha and hyper decoder hs networks that are used to obtain In this architecture the encoder subjects the input image x to ga, yielding the responses y with spatially varying standard deviations. The responses y are fed into ha, summarizing the distribution of standard deviations in z. z is then quantized compressed, and transmitted as side information. The encoder then uses the quantized vector to estimate σ, the spatial distribution of standard deviations, and uses it to compress and transmit the quantized image representation The decoder first recovers from the compressed signal. It then uses hs to obtain σ, which provides it with the correct probability estimates to successfully recover as well. It then feeds into gs to obtain the reconstructed image. When the hyper encoder and hyper decoder are added to the image compression network, the spatial redundancies of the quantized latent are reduced. The rightmost image in Fig. 2 correspond to the quantized latent when hyper encoder / decoder are used. Compared to middle right image, the spatial redundancies are significantly reduced, as the samples of the quantized latent are less correlated. As shown in Fig. 2, an image from the Kodak dataset is shown on the left. Visualization of the latent representation y of that image is shown on the middle left. In addition, standard deviations σ of the latent is shown on the middle right and latents y after the hyper prior (hyper encoder and decoder) network is introduced on the right. Fig. 3 illustrates a network architecture of an autoencoder implementing the hyperprior model. The left side shows an image autoencoder network, the right side corresponds to the hyperprior subnetwork. The analysis and synthesis transforms are denoted as ga and gs, respectively. Q represents quantization, and AE, AD represent arithmetic encoder and arithmetic decoder, respectively. The hyperprior model consists of two subnetworks, hyper encoder (denoted with ha) and hyper decoder (denoted with hs) . The hyper prior model generates a quantized hyper latent which comprises information about the probability distribution of the samples of the quantized latent is included in the bitstream and transmitted to the receiver (decoder) along with 2.3.4. Context Model Although the hyperprior model improves the modelling of the probability distribution of the quantized latent additional improvement can be obtained by utilizing an autoregressive model that predicts quantized latents from their causal context (Context Model) . The term auto-regressive means that the output of a process is later used as input to it. For example the context model subnetwork generates one sample of a latent, which is later used as input to obtain the next sample. Fig. 4 is a schematic diagram illustrating an example combined model configured to jointly optimize a context model along with a hyperprior and the autoencoder. The following Table illustrates meaning of different symbols. Table –Illustration of symbols A joint architecture was used where both hyperprior model subnetwork (hyper encoder and hyper decoder) and a context model subnetwork are utilized. The hyperprior and the context model are combined to learn a probabilistic model over quantized latents which is then used for entropy coding. As depicted in Fig. 4, the outputs of context subnetwork and hyper decoder subnetwork are combined by the subnetwork called Entropy Parameters, which generates the mean μ and scale (or variance) σ parameters for a Gaussian probability model. The gaussian probability model is then used to encode the samples of the quantized latents into bitstream with the help of the arithmetic encoder (AE) module. In the decoder the gaussian probability model is utilized to obtain the quantized latents from the bitstream by arithmetic decoder (AD) module. As shown in Fig. 4, the combined model jointly optimizes an autoregressive component that estimates the probability distributions of latents from their causal context (Context Model) along with a hyperprior and the underlying autoencoder. Real-valued latent representations are quantized (Q) to create quantized latents and quantized hyper-latents which are compressed into a bitstream using an arithmetic encoder (AE) and decompressed by an arithmetic decoder (AD) . The highlighted region corresponds to the components that are executed by the receiver (i.e. a decoder) to recover an image from a compressed bitstream. Typically, the latent samples are modeled as gaussian distribution or gaussian mixture models (not limited to) . According to Fig. 4, the context model and hyper prior are jointly used to estimate the probability distribution of the latent samples. Since a gaussian distribution can be defined by a mean and a variance (aka sigma or scale) , the joint model is used to estimate the mean and variance (denoted as μ and σ) . 2.3.5. The encoding process using joint auto-regressive hyper prior model Fig. 4 corresponds to the state-of-the-art compression method. In this section and the next, the encoding and decoding processes will be described separately. Fig. 5 depicts the encoding process. The input image is first processed with an encoder subnetwork. The encoder transforms the input image into a transformed representation called latent, denoted by y. y is then input to a quantizer block, denoted by Q, to obtained the quantized latent is then converted to a bitstream (bits1) using an arithmetic encoding module (denoted AE) . The arithmetic encoding block converts each sample of the into a bitstream (bits1) one by one, in a sequential order. The modules hyper encoder, context, hyper decoder, and entropy parameters subnetworks are used to estimate the probability distributions of the samples of the quantized latent The latent y is input to hyper encoder, which outputs the hyper latent (denoted by z) . The hyper latent is then quantized and a second bitstream (bits2) is generated using arithmetic encoding (AE) module. The factorized entropy module generates the probability distribution, that is used to encode the quantized hyper latent into bitstream. The quantized hyper latent includes information about the probability distribution of the quantized latent The Entropy Parameters subnetwork generates the probability distribution estimations, that are used to encode the quantized latent The information that is generated by the Entropy Parameters typically include a mean μand scale (or variance) σ parameters, that are together used to obtain a gaussian probability distribution. A gaussian distribution of a random variable x is defined as wherein the parameter μ is the mean or expectation of the distribution (and also its median and mode) , while the parameter σ is its standard deviation (or variance, or scale) . In order to define a gaussian distribution, the mean and the variance need to be determined. The entropy parameters module is used to estimate the mean and the variance values The subnetwork hyper decoder generates part of the information that is used by the entropy parameters subnetwork, the other part of the information is generated by the autoregressive module called context module. The context module generates information about the probability distribution of a sample of the quantized latent, using the samples that are already encoded by the arithmetic encoding (AE) module. The quantized latent is typically a matrix composed of many samples. The samples can be indicated using indices, such as or depending on the dimensions of the matrix The samples are encoded by AE one by one, typically using a raster scan order. In a raster scan order the rows of a matrix are processed from top to bottom, wherein the samples in a row are processed from left to right. In such a scenario (wherein the raster scan order is used by the AE to encode the samples into bitstream) , the context module generates the information pertaining to a sample using the samples encoded before, in raster scan order. The information generated by the context module and the hyper decoder are combined by the entropy parameters module to generate the probability distributions that are used to encode the quantized latent into bitstream (bits1) . Finally, the first and the second bitstream are transmitted to the decoder as result of the encoding process. It is noted that the other names can be used for the modules described above. In the above description, all of the elements in Fig. 5 are collectively called encoder. The analysis transform that converts the input image into latent representation is also called an encoder (or auto-encoder) . 2.3.6. The decoding process using joint auto-regressive hyper prior model Fig. 6 depicts the state-of-the-art decoding process. In the decoding process, the decoder first receives the first bitstream (bits1) and the second bitstream (bits2) that are generated by a corresponding encoder. The bits2 is first decoded by the arithmetic decoding (AD) module by utilizing the probability distributions generated by the factorized entropy subnetwork. The factorized entropy module typically generates the probability distributions using a predetermined template, for example using predetermined mean and variance values in the case of gaussian distribution. The output of the arithmetic decoding process of the bits2 is which is the quantized hyper latent. The AD process reverts to AE process that was applied in the encoder. The processes of AE and AD are lossless, meaning that the quantized hyper latent that was generated by the encoder can be reconstructed at the decoder without any change. After obtaining of it is processed by the hyper decoder, whose output is fed to entropy parameters module. The three subnetworks, context, hyper decoder and entropy parameters that are employed in the decoder are identical to the ones in the encoder. Therefore, the exact same probability distributions can be obtained in the decoder (as in encoder) , which is essential for reconstructing the quantized latent without any loss. As a result, the identical version of the quantized latent that was obtained in the encoder can be obtained in the decoder. After the probability distributions (e.g. the mean and variance parameters) are obtained by the entropy parameters subnetwork, the arithmetic decoding module decodes the samples of the quantized latent one by one from the bitstream bits1. From a practical standpoint, autoregressive model (the context model) is inherently serial, and therefore cannot be sped up using techniques such as parallelization. Finally, the fully reconstructed quantized latent is input to the synthesis transform (denoted as decoder in Fig. 6) module to obtain the reconstructed image. In the above description, all of the elements in Fig. 6 are collectively called decoder. The synthesis transform that converts the quantized latent into reconstructed image is also called a decoder (or auto-decoder) . 2.4. Multistage Context model for entropy probability modeling In 2.3.4, pixel CNN is employed as the context model to further build the relationship between different models, which provides better coding performance through the will designed context modeling. However, it limits the throughput of the probability prediction, because there is a strong dependence on the probability modeling of the elements, it needs to be decoded serially one by one, which might be a nightmare for high resolution images. To solve these issues, a multistage context model is proposed as the replacement of the context model, it provides better parallelism in GPU calculation and greatly reduces the decoding time of high-resolution images. Specifically, the Multistage Context model divided the latent information into 2 groups in channel dimension, as shown in the Fig. 7A. Fig. 7A illustrates a structure of the multistage context model. After splitting the channel dimension, for latent of each group, 4-stage context modeling will be applied to build context information for latent elements. The shape of the context modeling is provided in Fig. 7B. Fig. 7B illustrates a shape of the context model. Benefit from the design of the multistage context model, the times of the context prediction process is no longer dependent on the resolution of the input image, it comes to be a fixed number. In this design, it is only needed to perform the context prediction process 8 times, so that the decoding time is much faster than the original autoregressive context model. 2.5. Neural networks for video compression Similar to conventional video coding technologies, neural image compression serves as the foundation of intra compression in neural network-based video compression, thus development of neural network-based video compression technology comes later than neural network-based image compression but needs far more efforts to solve the challenges due to its complexity. Starting from 2017, a few researchers have been working on neural network-based video compression schemes. Compared with image compression, video compression needs efficient methods to remove inter-picture redundancy. Inter-picture prediction is then a crucial step in these works. Motion estimation and compensation is widely adopted but has not implemented by trained neural networks until recently. Studies on neural network-based video compression can be divided into two categories according to the targeted scenarios: random access and the low-latency. In random access case, it requires the decoding can be started from any point of the sequence, typically divides the entire sequence into multiple individual segments and each segment can be decoded independently. In low-latency case, it aims at reducing decoding time thereby usually merely temporally previous frames can be used as reference frames to decode subsequent frames. 2.5.1. Low-latency The early work first splits the video sequence frames into blocks and each block will choose one from two available modes, either intra coding or inter coding. If intra coding is selected, there is an associated auto-encoder to compress the block. If inter coding is selected, motion estimation and compensation are performed with tradition methods and a trained neural network will be used for residue compression. The outputs of auto-encoders are directly quantized and coded by the Huffman method. Another neural network-based video coding scheme with PixelMotionCNN was proposed. The frames are compressed in the temporal order, and each frame is split into blocks which are compressed in the raster scan order. Each frame will first be extrapolated with the preceding two reconstructed frames. When a block is to be compressed, the extrapolated frame along with the context of the current block are fed into the PixelMotionCNN to derive a latent representation. Then the residues are compressed by the variable rate image scheme. This scheme is performed on par with H. 264. Another end-to-end neural network-based video compression framework is then proposed, in which all the modules are implemented with neural networks. The scheme accepts current frame and the prior reconstructed frame as inputs and optical flow will be derived with a pre-trained neural network as the motion information. The motion information will be warped with the reference frame followed by a neural network generating the motion compensated frame. The residues and the motion information are compressed with two separate neural auto-encoders. The whole framework is trained with a single rate-distortion loss function. It achieves better performance than H. 264. An advanced neural network-based video compression scheme is proposed. It inherits and extends traditional video coding schemes with neural networks with the following major features: 1) using only one auto-encoder to compress motion information and residues; 2) motion compensation with multiple frames and multiple optical flows; 3) an on-line state is learned and propagated through the following frames over time. This scheme achieves better performance in MS-SSIM than HEVC reference software. An extended end-to-end neural network-based video compression framework is proposed afterwards. In this solution, multiple frames are used as references. It is thereby able to provide more accurate prediction of current frame by using multiple reference frames and associated motion information. In addition, motion field prediction is deployed to remove motion redundancy along temporal channel. Postprocessing networks are also introduced in this work to remove reconstruction artifacts from previous processes. The performance is better than H. 265 by a noticeable margin in terms of both PSNR and MS-SSIM. The scale-space flow is then proposed to replace commonly used optical flow by adding a scale parameter. It is reportedly achieving better performance than H. 264. A multi-resolution representation for optical flows is proposed. Concretely, the motion estimation network produces multiple optical flows with different resolutions and let the network to learn which one to choose under the loss function. The performance is better than H. 265. 2.5.2. Random access A frame interpolation based method was initially designed. The key frames are first compressed with a neural image compressor and the remaining frames are compressed in a hierarchical order. Motion compensation is performed in the perceptual domain, i.e. deriving the feature maps at multiple spatial scales of the original frame and using motion to warp the feature maps, which will be used for the image compressor. The method is reportedly on par with H. 264. Another interpolation-based video compression is then proposed, wherein the interpolation model combines motion information compression and image synthesis, and the same auto-encoder is used for image and residual. Afterwards, a neural network-based video compression method based on variational auto-encoders with a deterministic encoder is proposed. Concretely, the model consists of an auto-encoder and an auto-regressive prior. Different from previous methods, this method accepts a group of pictures (GOP) as inputs and incorporates a 3D autoregressive prior by taking into account of the temporal correlation while coding the laten representations. It provides comparative performance as H. 265. 2.5.3. Deep Contextual Video Compression The learning-based video compression model consists of an independent intra codec, a contextual or encoder EC, a contextual decoder DC, a motion encoder, a motion decoder, and an optical flow net. The intra codec is used to compress the key frame (I frame) in each GOP (group of images) . When compress the P frames xt, optical flow net is used to estimate the flow between previous decoded frame and current P frame xt. Flow map is employed as the motion vector. At encoder, a motion codec is used to compress the flow map and the reconstructed flow map is employed for consistency between encoding and decoding. The decoded frame is warped to context Ct based on the decoded optical flow When compressing P frame xt, xt is concatenated with context Ct and the result is fed to the contextual codec. The concatenation between xt and Ct is to let the network learn conditional coding. At decoder, the key frame or I frame is first decoded using the independent intra codec decoder. The optical flow is decoded using the motion codec decoder. The decoded frame is warped to context Ct based on the decoded optical flow. The inter codec is used to decompress the learned conditional coding information and it is concatenated with the context Ct to obtain the reconstructed frame The overall process is as follows: 2.5.4. Multi-Scale Deep Contextual Video Compression The compression of frame xt is taken as an example. Multi-scale features are first extracted from propagated feature The propagated feature is the feature extracted from previous frames. Motion vector is adopted to warp the multi-scale features to multi-scale local contexts The is concatenated to the current frame xt, is concatenated the to the mid-feature and the is concatenated to mid-feature The concatenation lets the network learn how to conduct conditional coding by itself. When decoding, multi-scale contexts are also concatenated to recover the frame. The overall process can be formulated as: where EC and ED are the contextual encoder and the contextual decoder. 2.6. Preliminaries Almost all the natural image / video is in digital format. A grayscale digital image can be represented by where is the set of values of a pixel, m is the image height and n is the image width. For example, is a common setting and in this case thus the pixel can be represented by an 8-bit integer. An uncompressed grayscale digital image has 8 bits-per-pixel (bpp) , while compressed bits are definitely less. A color image is typically represented in multiple channels to record the color information. For example, in the RGB color space an image can be denoted by with three separate channels storing Red, Green and Blue information. Similar to the 8-bit grayscale image, an uncompressed 8-bit RGB image has 24 bpp. Digital images / videos can be represented in different color spaces. The neural network-based video compression schemes are mostly developed in RGB color space while the traditional codecs typically use YUV color space to represent the video sequences. In YUV color space, an image is decomposed into three channels, namely Y, Cb and Cr, where Y is the luminance component and Cb / Cr are the chroma components. The benefits come from that Cb and Cr are typically down sampled to achieve pre-compression since human vision system is less sensitive to chroma components. A color video sequence is composed of multiple color images, called frames, to record scenes at different timestamps. For example, in the RGB color space, a color video can be denoted by X= {x0, x1, …, xt, …, xT-1} where T is the number of frames in this video sequence, If m=1080, n=1920, and the video has 50 frames-per-second (fps) , then the data rate of this uncompressed video is 1920×1080×8×3×50=2,488,320,000 bits-per-second (bps) , about 2.32 Gbps, which needs a lot storage thereby definitely needs to be compressed before transmission over the internet. Usually the lossless methods can achieve compression ratio of about 1.5 to 3 for natural images, which is clearly below requirement. Therefore, lossy compression is developed to achieve further compression ratio, but at the cost of incurred distortion. The distortion can be measured by calculating the average squared difference between the original image and the reconstructed image, i.e., mean-squared-error (MSE) . For a grayscale image, MSE can be calculated with the following equation. Accordingly, the quality of the reconstructed image compared with the original image can be measured by peak signal-to-noise ratio (PSNR) : where is the maximal value in e.g., 255 for 8-bit grayscale images. There are other quality evaluation metrics such as structural similarity (SSIM) and multi-scale SSIM (MS-SSIM) . To compare different lossless compression schemes, it is sufficient to compare either the compression ratio given the resulting rate or vice versa. However, to compare different lossy compression methods, it has to take into account both the rate and reconstructed quality. For example, to calculate the relative rates at several different quality levels, and then to average the rates, is a commonly adopted method; the average relative rate is known as Bjontegaard’s delta-rate (BD-rate) . There are other important aspects to evaluate image / video coding schemes, including encoding / decoding complexity, scalability, robustness, and so on. 3. Problems Neural video compression (NVC) has emerged as a promising solution for improving rate-distortion performance. However, previous neural video codecs primarily rely on convolutional neural networks (CNNs) to generate latent representations, which overlook the frequency characteristics of the features, resulting in redundancy in the latent space. Furthermore, to further improve the rate-distortion performance, variable bitrate strategy also needs to be well designed. 4. Detailed solutions 4.1. Details of the solutions The detailed solutions below should be considered as examples to explain general concepts. These solutions should not be interpreted in a narrow way. Furthermore, these solutions can be combined in any manner. Fig. 8A illustrates an example structure of a frequency decomposition transformer (FDT) module in accordance with some cases of the present disclosure, Fig. 8B illustrates a frequency (shifted) window-based self-attention (F(S) WSA) module of the FDT module in accordance with some cases of the present disclosure; Fig. 8C illustrates a frequency domain feed-forward network (FDFFN) of the FDT module in accordance with some cases of the present disclosure; and Fig. 8D illustrates a low frequency / high frequency- (shifted) window-based self-attention (LF / HF- (S) WSA) of the FDT module in accordance with some cases of the present disclosure. 1. In one example, frequency decomposition transformer (FDT) (Fig. 8A shows an example of the FDT module) will be utilized in current neural video compression. a) In one example, FDT will be employed in transformation module, i. In one example, FDT might be inserted after each residual block. ii. In one example, FDT may be applied after the convolution layer. iii. In one example, FDT could be utilized after activation function. b) In one example, FDT will be utilized in entropy module. i. In one example, FDT will be used after the convolution layers. ii. In one example, FDT may be applied in the entropy parameters network. iii. In one example, FDT will be utilized in checkerboard context model / autoregressive model. c) In one example, FDT might first taken tensor as the input and process it through convolution layer normalization and F (S) WSA module. i. In one example, there is a residual connection between the output of convolution layer and the output of F (S) WSA. ii. In one example, batch normalization might be utilized as the replacement of the layer normalization. iii. In one example, instance normalization might be utilized as the replacement of the layer normalization. iv. In one example, group normalization might be utilized as the replacement of the layer normalization. v. In one example, as it shows in the Fig. 8B, F (S) WSA might include single convolution and channel wise separation, which might process as high frequency information and low frequency information through LF / HF- (S) WSA and then combined through channel wise concatenation and convolution layer. 1. In one example, channel wise separation will be replaced with two individual convolution layers. 2. In one example, one or more layers will be added. 3. In one example, residual blocks might be applied inside this module. d) In one example, after F (S) WSA module, the output tensor might be processed by layer normalization and FDFFN. i. In one example, batch normalization might be utilized as the replacement of the layer normalization. ii. In one example, instance normalization might be utilized as the replacement of the layer normalization. iii. In one example, group normalization might be utilized as the replacement of the layer normalization. iv. In one example, as it shows in the Fig. 8C, FDFFN might include convolution layer, GELU, partition operation, FFT, weight factor multiplication, IFFT and merging operation. 1. In one example, convolution layer might be replaced with depth wise convolution. 2. In one example, GELU might be replaced with ReLU, LeakyReLU, ELU, SELU. e) In one example, after the process of F (S) WSA and FDFFN, a mirror network will be utilized as the inverse process. f) In one example, more FDFFN and F (S) WSA, might be applied inside the single FDT modules. 2. In one example, the synchronized variable rate strategy will be employed. a) In one example, for different training example, different λ will be applied in the loss function to enable the variable rate functionality, let R denotes rate, D represents distortion, loss function can be formulated as: L=R+λ×D i. In one example, λ can be different on batch size dimension, different sample might utilized different λ in the loss function. 1. In one example, λ can random picked from the finite λ set. a) In one example, λ set might be [85, 170, 380, 840] 2. In one example, λ can be an integer number that random picked form A to B. a) In one example, A might be 85 and B might be 840. ii. In one example, λ can be different on different GPU devices, each GPU handles a specific λ to reach a specific bitrate for synchronized optimization via backpropagation. b) In one example, the proposed synchronized variable rate strategy is enabled at finetuning stage. c) In one example, the proposed synchronized variable rate strategy is utilized at the beginning of the training. 4.2. Benefit of the solutions In accordance with some example implementations of the present disclosure, a Frequency Decomposition Transformer (FDT) and a variable bitrate strategy are employed on top of the existing end-to-end video compression model. By leveraging the FDT along with a synchronized training strategy, the proposed filters can significantly enhance coding performance. 5. Example Implementations Fig. 9 illustrates the proposed overall framework. In comparison to existing deep video compression frameworks, the Frequency Decomposition Transformer (FDT) is applied for the first time in the transformation module, which is positioned after each residual block (ResBlock) . Additionally, FDT is utilized in the entropy module: it is applied after the first convolution layers in the hyper encoder and before the last convolution layers in the hyper decoder.
[0048] More details of the cases of the present disclosure will be described below which are related to neural network (NN) -based visual data coding. As used herein, the term “visual data” may refer to a video, an image, a picture in a video, or any other visual data suitable to be coded.
[0049] Neural video compression (NVC) has emerged as a promising paradigm for improving rate-distortion performance. However, existing neural video codecs predominantly rely on convolutional neural networks ( (CNNs) with limited local receptive fields to generate the latent representations, often neglecting global-local spatial correlations. This leads to suboptimal feature modeling and redundancy in the latent space.
[0050] To solve the above problems and some other problems not mentioned, visual data processing solutions as described below are disclosed. The cases of the present disclosure should be considered as examples to explain the general concepts and should not be interpreted in a narrow way. Furthermore, these cases can be applied individually or combined in any manner.
[0051] Fig. 10 illustrates a flowchart of a method 1000 for visual data processing in accordance with cases of the present disclosure. At 1010, a conversion between visual data and a bitstream of the visual data is performed with a neural network (NN) -based model. In some cases, the conversion may include encoding the visual data into the bitstream. Alternatively or additionally, the conversion may include decoding the visual data from the bitstream. Alternatively or additionally, the conversion may include: generating the bitstream from the visual data, and the method 100 may further include: storing the bitstream in a non-transitory computer-readable recording medium.
[0052] As used herein, an NN-based model may be a model based on neural network technologies. For example, an NN-based model may specify sequence of neural network modules (also called architecture) and model parameters. The neural network module may comprise a set of neural network layers. Each neural network layer specifies a tensor operation which receives and outputs tensor, and each layer has trainable parameters. It should be understood that the possible implementations of the NN-based model described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any way.
[0053] In addition, the NN-based model includes at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation includes a feature representation associated with the visual data. By way of example rather than limitation, the NN-based module may be the frequency decomposition transformer (FDT) , which may also be referred to as dual-scale transformer (DST) . This will be described in detail below. As used herein, the term “feature representation” may represent an intermediate processing result within an NN-based model, and it also be referred to as “latent representation” .
[0054] In view of the above, the proposed method can advantageously jointly consider the local feature and the global feature associated with the feature representation that is associated with the visual data so as to reduce the redundancy in the feature representation space, and thus improving the coding efficiency.
[0055] In some cases, the local feature may correspond a high frequency component associated with the input feature representation, and the global feature may correspond to a low frequency component associated with the input feature representation. By way of example, the local feature may represent local details of the visual data, such as local texture details, texture variation within a small region, pixel gradients between adjacent pixels and a current pixel, and / or the like. The global feature may represent global information of the visual data, such as global structure information, statistical information (such as mean, variance or the like) of all pixels of the visual data. It should be understood that the above illustrations regarding the local and global features are described merely for purpose of description. The scope of the present disclosure is not limited in this respect.
[0056] In some cases, a first set of NN-based modules among the at least one NN-based module may be included in a transformation module of the NN-based model. Taking images as the example of the visual data, the transformation module may be used to convert the images in the pixel domain to feature representations in the latent space, or to convert feature representations in the latent space back to images in the pixel domain.
[0057] The position of the NN-based module in the transformation module of the NN-based model may be variously provided. By way of example rather than limitation, each residual block of the transformation module may be followed by one of the first set of NN-based modules. However, this is merely an example rather than a limitation. For example, some residual blocks of the transformation module may be followed by an NN-based module, while other residual blocks of the transformation module may not be followed by an NN-based module. Additionally or alternatively, a convolution layer of the transformation module may be followed by one of the first set of NN-based modules. Additionally or alternatively, an activation function of the transformation module may be followed by one of the first set of NN-based modules.
[0058] In some cases, a second set of NN-based modules among the at least one NN-based module may be included in an entropy module of the NN-based model. The position of the NN-based module in the entropy module of the NN-based model may be variously provided. By way of example rather than limitation, a convolution layer of the entropy module may be followed by one of the second set of NN-based modules. Additionally or alternatively, one of the second set of NN-based modules may be included in an entropy parameters network of the entropy module. Additionally or alternatively, one of the second set of NN-based modules may be included in a checkerboard context model or an autoregressive model of the entropy module.
[0059] It should be understood that the number and position of the NN-based module are described merely for purpose of description. For example, at each of the above positions, one or more NN-based module may be provided. In some examples, the one or more NN-based module may be provided at a position other than the above positions.
[0060] In some cases, each of the at least one NN-based module may include a first neural network, the first neural network may include a stack of a convolution layer, a first normalization layer and a first submodule, and the first submodule may be configured to determine the local feature and the global feature associated with the input feature representation. In some examples, the first submodule may be implemented as a global-local (shifted) window-based self-attention (GL (S) WSA) module, a.k.a., a frequency (shifted) window-based self-attention (F (S) WSA) module.
[0061] In some cases, the first normalization layer may include a layer normalization layer, a batch normalization layer, an instance normalization layer, a group normalization layer, or the like. Additionally or alternatively, there may be a residual connection between an output of the convolution layer and an output of the first submodule.
[0062] In some cases, an output of the first submodule may be generated by: obtaining a first feature representation and a second feature representation based on an input of the first submodule; determining the local feature from the first feature representation based on a window-based self-attention mechanism with a first window size; determining the global feature from the second feature representation based on a window-based self-attention mechanism with a second window size larger than the first window size; and combining the local feature and the global feature to obtain an output of the first submodule.
[0063] By way of example rather than limitation, the first feature representation and the second feature representation may be obtained through a single convolution layer and channel wise separation. Alternatively, the first feature representation and the second feature representation may be obtained through two individual convolution layers. Additionally or alternatively, the local feature and the global feature may be combined through channel wise concatenation and a convolution layer.
[0064] In some cases, the first submodule may include at least one of the following: one or more additional layers, or one or more residual blocks. The additional layers may include various layers, such as a convolution layer, an activation function, a normalization block, and / or the like. The additional layers may be disposed at any suitable positions within the first submodule.
[0065] By incorporating the first submodule, it is possible to effectively capture both global structure information and local texture details, so as to models local cues and global dependencies in a dual-scale manner, thereby reducing redundancy and improving rate–distortion (RD) performance, ultimately enabling more efficient video compression. It should be understood that the possible implementations of the first submodule described above are merely illustrative and therefore should not be construed as limiting the present disclosure in any way.
[0066] In some cases, the first neural network may further include a second normalization layer and a second submodule, and an output of the first submodule may be processed through the second normalization layer and the second submodule. For example, the second normalization layer may include a layer normalization layer, a batch normalization layer, an instance normalization layer, a group normalization layer, ort the like. In one example case, the second submodule may be implemented as frequency domain feed-forward network (FDFFN) as shown in Fig. 8C. Alternatively, the second submodule may be implemented as a cross-gated feed-forward network (CGFFN) as shown in Fig. 12C.
[0067] In one example case, the second submodule may include a stack of a convolution block, an activation function, a partition operation, a fast Fourier transform (FFT) block, a weight factor multiplication operation, an inverse fast Fourier transform (IFFT) block and a merging operation. In some examples, the activation function may include a Gaussian error linear unit (GELU) , a rectified linear unit (ReLU) , a leaky rectified linear unit (LeakyReLU) , an exponential linear unit (ELU) , a scaled exponential linear unit (SELU) , and / or the like. Additionally or alternatively, the convolution block may include a convolution layer or depth wise convolution.
[0068] In an alternative case, an output of the second submodule may be generated by: processing a feature representation inputted to the second submodule with the convolution block and the activation function; dividing the processed feature representation into a first channel feature part and a second channel feature part; generating a first modulated feature and a second modulated feature based on the first channel feature part and the second channel feature part; and generating the output of the second submodule based on the first modulated feature and the second modulated feature. By way of example rather than limitation, a feature representation inputted to the second submodule may be transformed, to obtain the processed feature representation. In some examples, the first channel feature part and the second channel feature part may also be referred to as a first channel group and a second channel group.
[0069] In some cases, generating the first modulated feature and the second modulated feature may include: generating a first gating weight based on the first channel feature part; generating a second gating weight based on the second channel feature part; generating the first modulated feature by applying the second gating weight to the first channel feature part; and generating the second modulated feature by applying the first gating weight to the second channel feature part. By way of example rather than limitation, the first gating weight may be generated by processing the first channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function, and the second gating weight may be generated by processing the second channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function. For example, the nonlinear activation function may include a GELU, and the nonlinear gating function may include a Sigmoid function.
[0070] Additionally or alternatively, generating the output of the second submodule may include: combining the first modulated feature and the second modulated feature to obtain a fused feature; applying a multi-layer perceptron (MLP) with an activation function to the fused feature to obtain a refined feature; and generating the output of the second submodule based on the refined fused feature. In some examples, the fused feature may also be referred to as a fused representation. In some examples, the MLP may be a two-layer perceptron. Additionally, the refined fused feature may be projected back to an original channel dimension to obtain the output of the second submodule. By way of example rather than limitation, the original channel dimension may be channel dimension of the input of the second submodule.
[0071] By incorporating the second submodule, it is possible to adaptively modulate complementary components, producing more compact and expressive latent representations. It should be understood that the possible implementations of the second submodule described above are merely illustrative and therefore should not be construed as limiting the present disclosure in any way.
[0072] In some cases, each of the at least one NN-based module may further include a second neural network configured to perform an inverse process of the first neural network. By way of example rather than limitation, the second neural network may be a mirror network of the first neural network. In some cases, each of the at least one NN-based module may include more than one first submodule and more than one second submodule.
[0073] In some cases, during a training process of the NN-based model, the NN-based model may be trained in a distributed manner at a plurality of devices, each of the plurality of devices may be assigned to a distinct bitrate and determines respective gradients, the gradients of the plurality of devices may be synchronized and aggregated for updating parameters of the NN-based model. For example, each device may be assigned a bitrate, and bitrates assigned to the plurality of devices may be different from each other. By way of example rather than limitation, the devices may include graphics processing unit (GPU) device, a central processing unit (CPU) device, and / or the like.
[0074] In some cases, at each of the plurality of devices, a loss function for training the NN-based model may be as follows: L=R+λ×D, where L represents to a training loss, R represents to bitrate, D represents a distortion, and λ represents a weighting factor. By way of example rather than limitation, the loss in the loss function may be rate-distortion loss. By way of example rather than limitation, bitrate represented by R may also be referred to as rate, or the number of bits in time t.
[0075] In some cases, values of the weighting factor λ may be different for different training examples. By adjusting the weighting factor λ, it is possible to adjust the bitrate, so as to enable the variable rate functionality. Additionally or alternatively, values of the weighting factor λ may be different for different samples on a batch size dimension. Additionally or alternatively, a value of the weighting factor λ may be determined from a set of candidate values. For example, the set of candidate values may include 85, 170, 380, and 840, or the set of candidate values may include all integers within a value range from a first integer to a second integer. In some examples, the first integer may be 85, and the second integer may be 840. It should be understood that the specific values recited herein are intended to be examples rather than limiting the scope of the present disclosure.
[0076] In some cases, values of the weighting factor λ may be different for different devices of the plurality of devices. In some cases, the above-described training process may be applied at a finetuning stage of the NN-based model, or the training process may be applied from the beginning of training the NN-based model.
[0077] In aid of the above training process, a synchronization mechanism that leverages multi-devices parallel training, with each device dedicated to a specific bitrate and synchronized via gradient backpropagation for joint optimization, may be provided. Thus, it is possible to achieve the higher coding performance compared to the existing solutions, and to significantly improve the coding performance.
[0078] For purpose of illustration, the proposed solution will be discussed in more detail according to a non-limiting example below.
[0079] Fig. 11 illustrates an overview of the neural video codec based on the proposed DST-based contextual encoder-decoder module. AE and AD denote arithmetic encoder and arithmetic decoder, respectively. The example architecture shown in Fig. 11 builds on the deep contextual video compression with diverse contexts (DCVC-DC) , but focuses more on exploiting local and global dependencies of the nonlinear transformation through window-based self-attention. The workflow of overall framework is illustrated in Fig. 11.
[0080] Architecture of the DST-based contextual encoder-decoder and associated hyperprior encoder-decoder has been shown in Fig. 9. xt, are input frame and reconstructed frame. are the multi-scale temporal contexts. is propagated feature for the subsequent encoding. qenc and qdec are the quantization steps (QS) adopted from DCVC-DC. yt is the latent representation and zt is the hyperprior latent representation. Q represents the quantization. μt and σt of are the estimated means and scales of quantized latent representation yt by the entropy model for AE and AD.
[0081] For motion estimation, the input frame xt and the reference frame (previously decoded frame) are fed into the motion estimation module to generate the motion vector vt, represented by optical flows.
[0082] The estimated motion vector vt is encoded into a latent representation mt by the motion encoder, which is then quantized and converted into a binary bitstream using arithmetic encoder (AE) . The arithmetic decoder (AD) then reconstructs from the transmitted bitstream, and the moption decoder inversely transforms it back into the reconstructed motion vector Alearnable motion quantization step and its inverse counterpart are embedded in the motion encoder and decoder, respectively, to support bitrate adjustment within a single model.
[0083] A temporal contextual mining module is applied to perform feature-domain motion compensation on the corresponding reconstructed motion vector and the propagated reference feature thereby producing multi-scale temporal contexts
[0084] The architecture of DST-based contextual encoder-decoder may refer to Fig. 9. Given the input frame xt, the largest temporal context is concatenated and fed into the contextual ender. During the transformation from the pixel-wise frame xt, to the latent representation yt, additional temporal contexts and are concatenated at other scales and passed through the encoder. Similarly, learned multi-scale temporal contexts are used as conditions in the decoder. Once the compact latent representation yt is obtained, it is further processed by the hyperprior encoder to generate zt. After quantization, the quantized zt is modeled by a fully factorized density model to estimate its probability distribution, which is then compressed into a bitstream using arithmetic encoding (AE) .
[0085] At the decoder side, the bitstream is entropy-decoded by the arithmetic decoder (AD) to reconstruct which is subsequently transformed by the hyperprior decoder into side information. This side information guides the entropy model to estimate the means μt and scales σt of The quantized is then entropy-coded via AE, and at the decoder side it is recovered by AD and inversely transformed through the context decoder and reconstruction module to obtain the decoded frame Learnable quantization steps qenc and their inverse counterparts qdec are embedded in the contextual encoder and decoder, respectively, to enable variable bitrate adaptation.
[0086] Considering that the local receptive field of contextual encoder and decoder, a Dual-Scale Transformer (DST) block is proposed to better capture both local texture details and global structure information, thereby enhancing the nonlinear transformation capability. Moreover, DCVC family rely on asynchronous variable bitrate training that sequentially optimizes different bitrates in a suboptimal manner. A Variable Bitrate Synchronization (VBRS) strategy is introduced to achieve collaborative optimization across multiple bitrates within a single model. More details about DST and VBRS will be introduced later.
[0087] A frame generation is applied, which ties two U-Nets to reconstruct the decoded features into the pixel-domain reconstructed frame and feature-domain representation and is further propagated as the reference feature for encoding the subsequent frame xt+1.
[0088] The latent representation of quantized motion vector and contextual latent representation are assumed to follow Laplacian distributions. The parameters of Laplacian distributions are estimated through a combination of the hyperprior, quadtree-based spatial, and conditional temporal entropy models. Entropy coding and Decoding are then performed with an arithmetic encoder and decoder, respectively.
[0089] The Dual-Scale Transformer Block is shown in Figs. 12A to 12D. Fig. 12A illustrates an overview of a Dual-Scale Transformer (DST) block; Fig. 12B illustrates a Global-Local (Shifted) Window-based Self-Attention (GL (S) WSA module of the DST block; Fig. 12C illustrates a Cross-Gated Feed-Forward Network (CGFFN) of the DST block; and Fig. 12D illustrates a Global / Local (Shifted) Window based Self-Attention (Global / Local (S) WSA) of the DST block;
[0090] As illustrated in Fig. 12A, the proposed Dual-Scale Transformer (DST) block incorporates a Global-Local (Shifted) Window-based Self-Attention (GL (S) WSA) mechanism, which decomposes the encoded features into local texture details and global structure information. Subsequently, a Cross-Gated Feed-Forward Network (CGFFN) adaptively modulates these components, thereby enhancing representation efficiency and reducing latent redundancy. In addition, a feature padding operation is adopted to seamlessly handle images of arbitrary resolutions.
[0091] The self-attention (SA) mechanism of transformer inherently functions as a low-pass filter, with smaller local windows capturing texture details and larger local windows modeling structural information. To balance computational efficiency and representational capacity, it is proposed a Global-Local (Shifted) Window based Self-Attention (GL (S) WSA) module which applies self-attention with two different window sizes to capture both local and global dependencies. Specifically, the window sizes are set to 4k ×4k and k ×k respectively, where k denotes the basic window size.
[0092] The smaller k ×k windows emphasize high-frequency details, while the larger 4k ×4k windows capture low-frequency structural patterns. The self-attention computed with these two window shapes is referred to as Local (S) WSA and Global (S) WSA, respectively.
[0093] As depicted in Fig. 12B, given the input feature representation where C represents the number of channels, H denotes the height, and W indicates the width, the input may be first linearly projected into N Heads. Then, these Heads are evenly divided into two parallel groups, which are designated to perform Local (S) WSA and Global (S) WSA, as shown in Fig. 12D. For global dependencies, X is padded to 4kx feature automatically. Then padded feature is uniformly divided into non-overlapping windows where each window contains 4k ×4k tokens. Then, the output of the Global (S) WSA for the n-th Head can be formulated as follows. where Q, K, V refer to the projection matrices of query, key and value for n-th Head. Finally, the global structure information is cropped back to original size.
[0094] Similarly, the output of Local (Shifted) Window Self-Attention for n-th Head are denoted as Local (S) WSAn (X) . Finally, two parallel groups can be concatenated to form the output of GL (S) WSA. GL(S) WSA (X) =Concat [Head1, …, Headn] O where O is a projection matrix designed to facilitate interactions between the texture detail and structure information.
[0095] To enhance cross-channel interaction and adaptively balance complementary feature subspaces, a Cross-Gated Feed Forward Network (CGFFN) is designed. As illustrated in Fig. 12C, given the hidden representation Y, it is first feed into a convolutional layer followed Iby the GELU activation to obtain the transformed feature Z. Subsequently, Z is divided into two channel groups: [ZA, ZB] = Split (Z) (3)
[0096] Two lightweight gating branches are then employed to enable mutual modulation between the two subspaces: GA = Sigmoid (Linear (GELU (Linear (ZA) ) ) ) GB = Sigmoid (Linear (GELU (Linear (ZB) ) ) ) (4) where GELU (·) and Sigmoid (·) denote the nonlinear activation and gating functions, respectively. The obtained gating weights GA and GB adaptively control the information exchange between the two channel groups: and the modulated features are concatenated to form the fused representation:
[0097] A two-layer MLP with GELU activation is then applied to refine the fused feature:
[0098] Finally, F is projected back to the original channel dimension to produce output This cross-gated mechanism allows each subspace to dynamically filter and enhance the complementary cues from its counterpart, effectively capturing complex dependencies such as global and local relations within a single FFN layer, while maintaining a lightweight computational footprint.
[0099] The variable bitrate synchronization strategy will be discussed with reference to Fig. 13A and Fig. 13B. Fig. 13A illustrates an asynchronous training strategy on multiple GPU; and Fig. 13B illustrates a synchronous training strategy on multiple GPU.
[0100] Variable bitrate training remains a critical challenge in neural video compression (NVC) , as it often results in performance degradation. Existing solution typically adjust bitrate by varying the quantization step (QS) , thereby allowing smooth bitrate control during training. However, these asynchronous strategies are suboptimal. In particular, as illustrated in Fig. 13A, in the existing solution, each bitrate is trained sequentially without exploiting the potential correlations across different bitrate settings.
[0101] Specifically, VBRS employs multi-GPU parallel training, where each GPU is assigned to a specific bitrate, and synchronous gradient backpropagation ensures joint optimization across all bitrates within each training iteration. This design maintains consistent training progress among different bitrates, enabling efficient dynamic bit allocation and ultimately improving compression performance.
[0102] Although smooth bitrate adjustment has been implemented in a single model, an asynchronous training strategy that results in suboptimal bit allocation and rate-distortion (RD) performance, is used. To solve the above problems and some other problems not mentioned, the Variable Bitrate Synchronization (VBRS) strategy is proposed, as illustrated in Fig. 13B.
[0103] It is first provided theoretical analysis for why synchronous training under a distributed multi-GPU setting with the Adam optimizer achieves superior compression performance compared to asynchronous training.
[0104] Let denote the model parameters, and let Li (θ) represent the loss function associated with bitrate i∈ {1, 2, …, n } . The joint training objective can be expressed as a weight sum: where αi≥0 and are the weighting coefficients.
[0105] In the synchronous training setting, each GPU is assigned to a distinct bite and computes its respective gradient The gradients are synchronized and aggregated as:
[0106] These aggregated gradients are then used in a unified Adam update with shared moment estimates: where mt and vt are the bias-corrected estimates of the firrst and second moments. Here, β1= 0.9, β2 = 0.999, η is the learning rate, and ensures numerical stability.
[0107] In contrast, asynchronous training allocates mixed-bitrate batches to each GPU. Consequently, the stochastic gradient computed by each GPU corresponds to a weighted subset of bitrate configurations: where B denotes the sampled set of bitrates, b indexes the sample distances, and ∈t represents sampling noise arising from limited and imbalanced bitrate instances per batch.
[0108] This introduces three key issues: (1) increased gradient variance, (2) inconsistent moment estimates in Adam, and (3) unstable approximation of the true joint objective. By contrast, the synchronous GPU-per-bitrate design preserves gradient directions aligned with the true objective, reduces gradient noise, and ensures consistent momentum accumulation through synchronized Adam updates. This alignment leads to more stable convergence and lower final BD-rate.
[0109] Formally, let θ* denote the minimizer of L (θ) . The additional stochasticity in asynchronous updates biases the parameter trajectory: which degrades compression efficiency across multiple bitrates.
[0110] It is further set up a distributed training configuration with n GPUs, where each GPU is dedicated to a quantization step corresponding to a specific bitrate. The quantization steps are denoted as {q1, q2, …, qn} , with qi used on the i-th GPU. Each GPU quantizes its input latent feature using where yi is the latent representation for bitrate i.
[0111] A rate-distortion loss is defined as follows: where T is the length of the training sequence, Dt and Rt denote the distortion and the number of bits in time t, respectively, and A balance the bitrate and distortion.
[0112] The forward pass is executed independently on each GPU. During backpropagation, gradients are computed as and then synchronized, ensuring consistent parameter updates across all GPUs. This allows the model to converge towards an optimal solution that improves compression efficiency simultaneously at multiple bitrates.
[0113] In a nutshell, in order to generate the compact latent representation, there is provided the Dual-Scale Transformer (DST) block that utilizes different window shapes to capture both global structure information and local texture detail, while adaptively modulating these components through a Cross-Gated Feed-Forward Network (CGFFN) to achieve more compact and expressive representations. Furthermore, to overcome the limitations of the asynchronous training, a Variable Bitrate Synchronization (VBRS) strategy is introduced, which enables stable convergence and optimal coordination across multiple GPUs in distributed training. With the combination of the DST block and VBRS strategy, the above-described neural video codec achieves superior performance compared with previous state-of-the-art NVC methods as well as the traditional non-NN-based coding software.
[0114] In view of the above, the solutions in accordance with some cases of the present disclosure can advantageously improve coding efficiency and coding quality.
[0115] According to further cases of the present disclosure, an apparatus for visual data processing is provided. The apparatus includes a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform any of the above methods.
[0116] According to further cases of the present disclosure, a non-transitory computer-readable storage medium is provided. The non-transitory computer-readable storage medium stores instructions that cause a processor to perform any of the above methods.
[0117] According to further cases of the present disclosure, a non-transitory computer-readable recording medium is provided. The non-transitory computer-readable recording medium stores a bitstream of visual data which is generated by a method performed by an apparatus for visual data processing. The method includes: generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model includes at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation includes a feature representation associated with the visual data.
[0118] According to further cases of the present disclosure, a method for storing a bitstream of visual data is provided. The method includes: generating the bitstream of the visual data with a neural network (NN) -based model. The NN-based model includes at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation includes a feature representation associated with the visual data; and storing the bitstream in a non-transitory computer-readable recording medium.
[0119] According to further cases of the present disclosure, a computer program is provided. The computer program is for performing any of the above methods when the computer program runs on a computer.
[0120] According to further cases of the present disclosure, a computer program product is provided. The computer program product is tangibly stored in a computer storage medium and comprises computer-executable instructions that, when executed by a device, cause the device to perform any of the above methods.
[0121] Implementations of the present disclosure can be described in view of the following clauses, the features of which can be combined in any reasonable manner.
[0122] Clause 1. A method for visual data processing, comprising: performing a conversion between visual data and a bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.
[0123] Clause 2. The method of clause 1, wherein the local feature corresponds a high frequency component associated with the input feature representation, and the global feature corresponds to a low frequency component associated with the input feature representation.
[0124] Clause 3. The method of any of clauses 1-2, wherein a first set of NN-based modules among the at least one NN-based module is comprised in a transformation module of the NN-based model.
[0125] Clause 4. The method of clause 3, wherein each residual block of the transformation module is followed by one of the first set of NN-based modules.
[0126] Clause 5. The method of any of clauses 3-4, wherein a convolution layer of the transformation module is followed by one of the first set of NN-based modules.
[0127] Clause 6. The method of any of clauses 3-5, wherein an activation function of the transformation module is followed by one of the first set of NN-based modules.
[0128] Clause 7. The method of any of clauses 1-6, wherein a second set of NN-based modules among the at least one NN-based module is comprised in an entropy module of the NN-based model.
[0129] Clause 8. The method of clause 7, wherein a convolution layer of the entropy module is followed by one of the second set of NN-based modules.
[0130] Clause 9. The method of any of clauses 7-8, wherein one of the second set of NN-based modules is comprised in an entropy parameters network of the entropy module.
[0131] Clause 10. The method of any of clauses 7-9, wherein one of the second set of NN-based modules is comprised in a checkerboard context model or an autoregressive model of the entropy module.
[0132] Clause 11. The method of any of clauses 1-10, wherein each of the at least one NN-based module comprises a first neural network, the first neural network comprises a stack of a convolution layer, a first normalization layer and a first submodule, and the first submodule is configured to determine the local feature and the global feature associated with the input feature representation.
[0133] Clause 12. The method of clause 11, wherein the first normalization layer comprises one of the following: a layer normalization layer, a batch normalization layer, an instance normalization layer, or a group normalization layer.
[0134] Clause 13. The method of any of claim 11-12, wherein there is a residual connection between an output of the convolution layer and an output of the first submodule.
[0135] Clause 14. The method of any of clauses 11-13, wherein an output of the first submodule is generated by: obtaining a first feature representation and a second feature representation based on an input of the first submodule; determining the local feature from the first feature representation based on a window-based self-attention mechanism with a first window size; determining the global feature from the second feature representation based on a window-based self-attention mechanism with a second window size larger than the first window size; and combining the local feature and the global feature to obtain an output of the first submodule.
[0136] Clause 15. The method of clause 14, wherein the first feature representation and the second feature representation are obtained through a single convolution layer and channel wise separation, or the first feature representation and the second feature representation is obtained through two individual convolution layers.
[0137] Clause 16. The method of any of clauses 14-15, wherein the local feature and the global feature is combined through channel wise concatenation and a convolution layer.
[0138] Clause 17. The method of any of clauses 14-16, wherein the first submodule comprises at least one of the following: one or more additional layers, or one or more residual blocks.
[0139] Clause 18. The method of any of clauses 11-17, wherein the first neural network further comprises a second normalization layer and a second submodule, and an output of the first submodule is processed through the second normalization layer and the second submodule.
[0140] Clause 19. The method of clause 18, wherein the second normalization layer comprises one of the following: a layer normalization layer, a batch normalization layer, an instance normalization layer, or a group normalization layer.
[0141] Clause 20. The method of any of clauses 18-19, wherein the second submodule comprises a stack of a convolution block, an activation function, a partition operation, a fast Fourier transform (FFT) block, a weight factor multiplication operation, an inverse fast Fourier transform (IFFT) block and a merging operation.
[0142] Clause 21. The method of clause 20, wherein the convolution block comprises a convolution layer or depth wise convolution.
[0143] Clause 22. The method of any of clauses 20-21, wherein the activation function comprises one of the following: a Gaussian error linear unit (GELU) , a rectified linear unit (ReLU) , a leaky rectified linear unit (LeakyReLU) , an exponential linear unit (ELU) , a scaled exponential linear unit (SELU) .
[0144] Clause 23. The method of any of clauses 20-22, wherein an output of the second submodule is generated by: processing a feature representation inputted to the second submodule with the convolution block and the activation function; dividing the processed feature representation into a first channel feature part and a second channel feature part; generating a first modulated feature and a second modulated feature based on the first channel feature part and the second channel feature part; and generating the output of the second submodule based on the first modulated feature and the second modulated feature.
[0145] Clause 24. The method of clause 23, wherein generating the first modulated feature and the second modulated feature comprises: generating a first gating weight based on the first channel feature part; generating a second gating weight based on the second channel feature part; generating the first modulated feature by applying the second gating weight to the first channel feature part; and generating the second modulated feature by applying the first gating weight to the second channel feature part.
[0146] Clause 25. The method of clause 24, wherein the first gating weight is generated by processing the first channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function, and the second gating weight is generated by processing the second channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function.
[0147] Clause 26. The method of clause 25, wherein the nonlinear activation function comprises a GELU, and the nonlinear gating function comprises a Sigmoid function.
[0148] Clause 27. The method of any of clauses 23-26, wherein generating the output of the second submodule comprises: combining the first modulated feature and the second modulated feature to obtain a fused feature; applying a multi-layer perceptron (MLP) with an activation function to the fused feature to obtain a refined feature; and generating the output of the second submodule based on the refined fused feature.
[0149] Clause 28. The method of clause 27, wherein the refined fused feature is projected back to an original channel dimension to obtain the output of the second submodule.
[0150] Clause 29. The method of any of clauses 11-28, wherein each of the at least one NN-based module further comprises a second neural network configured to perform an inverse process of the first neural network.
[0151] Clause 30. The method of any of clauses 18-29, wherein each of the at least one NN-based module comprises more than one first submodule and more than one second submodule.
[0152] Clause 31. The method of any of clauses 1-30, wherein during a training process of the NN-based model, the NN-based model is trained in a distributed manner at a plurality of devices, each of the plurality of devices is assigned to a distinct bitrate and determines respective gradients, the gradients of the plurality of devices are synchronized and aggregated for updating parameters of the NN-based model.
[0153] Clause 32. The method of clause 31, where at each of the plurality of devices, a loss function for training the NN-based model is as follows: L=R+λ×D, wherein L represents to a training loss, R represents to bitrate, D represents a distortion, and λ represents a weighting factor.
[0154] Clause 33. The method of clause 32, wherein values of the weighting factor λ are different for different training examples.
[0155] Clause 34. The method of any of clauses 32-33, wherein values of the weighting factor λ are different for different samples on a batch size dimension.
[0156] Clause 35. The method of any of clauses 32-34, wherein a value of the weighting factor λ is determined from a set of candidate values.
[0157] Clause 36. The method of clause 35, wherein the set of candidate values comprises 85, 170, 380, and 840, or wherein the set of candidate values comprises all integers within a value range from a first integer to a second integer.
[0158] Clause 37. The method of clause 36, wherein the first integer is 85, and the second integer is 840.
[0159] Clause 38. The method of clause 32, wherein values of the weighting factor λ are different for different devices of the plurality of devices.
[0160] Clause 39. The method of any of clauses 31-38, wherein the training process is applied at a finetuning stage of the NN-based model, or the training process is applied from the beginning of training the NN-based model.
[0161] Clause 40. The method of any of clauses 1-39, wherein the visual data comprise a video, a picture of the video, or an image.
[0162] Clause 41. The method of any of clauses 1-40, wherein the conversion comprises encoding the visual data into the bitstream.
[0163] Clause 42. The method of any of clauses 1-40, wherein the conversion comprises decoding the visual data from the bitstream.
[0164] Clause 43. The method of any of clauses 1-40, wherein the conversion comprises: generating the bitstream from the visual data, and the method further comprises: storing the bitstream in a non-transitory computer-readable recording medium.
[0165] Clause 44. An apparatus for visual data processing comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform a method in accordance with any of clauses 1-43.
[0166] Clause 45. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of clauses 1-43.
[0167] Clause 46. A non-transitory computer-readable recording medium storing a bitstream of visual data which is generated by a method performed by an apparatus for visual data processing, wherein the method comprises: generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.
[0168] Clause 47. A method for storing a bitstream of visual data, comprising: generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data; and storing the bitstream in a non-transitory computer-readable recording medium.
[0169] Clause 48. A computer program for performing a method in accordance with any of clauses 1-43 when the computer program runs on a computer.
[0170] Clause 49. A computer program product tangibly stored in a computer storage medium and comprising computer-executable instructions that, when executed by a device, cause the device to perform a method in accordance with any of clauses 1-43. Example Device
[0171] Fig. 14 illustrates a block diagram of a computing device 1400 in which various cases of the present disclosure can be implemented. The computing device 1400 may be implemented as or included in the source device 110 (or the visual data encoder 114 or 200) or the destination device 120 (or the visual data decoder 124 or 300) .
[0172] It would be appreciated that the computing device 1400 shown in Fig. 14 is merely for purpose of illustration, without suggesting any limitation to the functions and scopes of the cases of the present disclosure in any manner.
[0173] As shown in Fig. 14, the computing device 1400 includes a general-purpose computing device 1400. The computing device 1400 may at least comprise one or more processors or processing units 1410, a memory 1420, a storage unit 1430, one or more communication units 1440, one or more input devices 1450, and one or more output devices 1460.
[0174] In some cases, the computing device 1400 may be implemented as any user terminal or server terminal having the computing capability. The server terminal may be a server, a large-scale computing device or the like that is provided by a service provider. The user terminal may for example be any type of mobile terminal, fixed terminal, or portable terminal, including a mobile phone, station, unit, device, multimedia computer, multimedia tablet, Internet node, communicator, desktop computer, laptop computer, notebook computer, netbook computer, tablet computer, personal communication system (PCS) device, personal navigation device, personal digital assistant (PDA) , audio / video player, digital camera / video camera, positioning device, television receiver, radio broadcast receiver, E-book device, gaming device, or any combination thereof, including the accessories and peripherals of these devices, or any combination thereof. It would be contemplated that the computing device 1400 can support any type of interface to a user (such as “wearable” circuitry and the like) .
[0175] The processing unit 1410 may be a physical or virtual processor and can implement various processes based on programs stored in the memory 1420. In a multi-processor system, multiple processing units execute computer executable instructions in parallel so as to improve the parallel processing capability of the computing device 1400. The processing unit 1410 may also be referred to as a central processing unit (CPU) , a microprocessor, a controller or a microcontroller.
[0176] The computing device 1400 typically includes various computer storage medium. Such medium can be any medium accessible by the computing device 1400, including, but not limited to, volatile and non-volatile medium, or detachable and non-detachable medium. The memory 1420 can be a volatile memory (for example, a register, cache, Random Access Memory (RAM) ) , a non-volatile memory (such as a Read-Only Memory (ROM) , Electrically Erasable Programmable Read-Only Memory (EEPROM) , or a flash memory) , or any combination thereof. The storage unit 1430 may be any detachable or non-detachable medium and may include a machine-readable medium such as a memory, flash memory drive, magnetic disk or another other media, which can be used for storing information and / or data and can be accessed in the computing device 1400.
[0177] The computing device 1400 may further include additional detachable / non-detachable, volatile / non-volatile memory medium. Although not shown in Fig. 14, it is possible to provide a magnetic disk drive for reading from and / or writing into a detachable and non-volatile magnetic disk and an optical disk drive for reading from and / or writing into a detachable non-volatile optical disk. In such cases, each drive may be connected to a bus (not shown) via one or more data medium interfaces.
[0178] The communication unit 1440 communicates with a further computing device via the communication medium. In addition, the functions of the components in the computing device 1400 can be implemented by a single computing cluster or multiple computing machines that can communicate via communication connections. Therefore, the computing device 1400 can operate in a networked environment using a logical connection with one or more other servers, networked personal computers (PCs) or further general network nodes.
[0179] The input device 1450 may be one or more of a variety of input devices, such as a mouse, keyboard, tracking ball, voice-input device, and the like. The output device 1460 may be one or more of a variety of output devices, such as a display, loudspeaker, printer, and the like. By means of the communication unit 1440, the computing device 1400 can further communicate with one or more external devices (not shown) such as the storage devices and display device, with one or more devices enabling the user to interact with the computing device 1400, or any devices (such as a network card, a modem and the like) enabling the computing device 1400 to communicate with one or more other computing devices, if required. Such communication can be performed via input / output (I / O) interfaces (not shown) .
[0180] In some cases, instead of being integrated in a single device, some or all components of the computing device 1400 may also be arranged in cloud computing architecture. In the cloud computing architecture, the components may be provided remotely and work together to implement the functionalities described in the present disclosure. In some cases, cloud computing provides computing, software, data access and storage service, which will not require end users to be aware of the physical locations or configurations of the systems or hardware providing these services. In various cases, the cloud computing provides the services via a wide area network (such as Internet) using suitable protocols. For example, a cloud computing provider provides applications over the wide area network, which can be accessed through a web browser or any other computing components. The software or components of the cloud computing architecture and corresponding data may be stored on a server at a remote position. The computing resources in the cloud computing environment may be merged or distributed at locations in a remote data center. Cloud computing infrastructures may provide the services through a shared data center, though they behave as a single access point for the users. Therefore, the cloud computing architectures may be used to provide the components and functionalities described herein from a service provider at a remote location. Alternatively, they may be provided from a conventional server or installed directly or otherwise on a client device.
[0181] The computing device 1400 may be used to implement visual data encoding / decoding in cases of the present disclosure. The memory 1420 may include one or more visual data coding modules 1425 having one or more program instructions. These modules are accessible and executable by the processing unit 1410 to perform the functionalities of the various cases described herein.
[0182] In the example cases of performing visual data encoding, the input device 1450 may receive visual data as an input 1470 to be encoded. The visual data may be processed, for example, by the visual data coding module 1425, to generate an encoded bitstream. The encoded bitstream may be provided via the output device 1460 as an output 1480.
[0183] In the example cases of performing visual data decoding, the input device 1450 may receive an encoded bitstream as the input 1470. The encoded bitstream may be processed, for example, by the visual data coding module 1425, to generate decoded visual data. The decoded visual data may be provided via the output device 1460 as the output 1480.
[0184] While this disclosure has been particularly shown and described with references to example cases thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present application as defined by the appended claims. Such variations are intended to be covered by the scope of this present application. As such, the foregoing description of cases of the present application is not intended to be limiting.
Claims
1.A method for visual data processing, comprising:performing a conversion between visual data and a bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.2.The method of claim 1, wherein the local feature corresponds a high frequency component associated with the input feature representation, and the global feature corresponds to a low frequency component associated with the input feature representation.3.The method of any of claims 1-2, wherein a first set of NN-based modules among the at least one NN-based module is comprised in a transformation module of the NN-based model.4.The method of claim 3, wherein each residual block of the transformation module is followed by one of the first set of NN-based modules.5.The method of any of claims 3-4, wherein a convolution layer of the transformation module is followed by one of the first set of NN-based modules.6.The method of any of claims 3-5, wherein an activation function of the transformation module is followed by one of the first set of NN-based modules.7.The method of any of claims 1-6, wherein a second set of NN-based modules among the at least one NN-based module is comprised in an entropy module of the NN-based model.8.The method of claim 7, wherein a convolution layer of the entropy module is followed by one of the second set of NN-based modules.9.The method of any of claims 7-8, wherein one of the second set of NN-based modules is comprised in an entropy parameters network of the entropy module.10.The method of any of claims 7-9, wherein one of the second set of NN-based modules is comprised in a checkerboard context model or an autoregressive model of the entropy module.11.The method of any of claims 1-10, wherein each of the at least one NN-based module comprises a first neural network, the first neural network comprises a stack of a convolution layer, a first normalization layer and a first submodule, and the first submodule is configured to determine the local feature and the global feature associated with the input feature representation.12.The method of claim 11, wherein the first normalization layer comprises one of the following: a layer normalization layer, a batch normalization layer, an instance normalization layer, or a group normalization layer.13.The method of any of claim 11-12, wherein there is a residual connection between an output of the convolution layer and an output of the first submodule.14.The method of any of claims 11-13, wherein an output of the first submodule is generated by:obtaining a first feature representation and a second feature representation based on an input of the first submodule;determining the local feature from the first feature representation based on a window-based self-attention mechanism with a first window size;determining the global feature from the second feature representation based on a window-based self-attention mechanism with a second window size larger than the first window size; andcombining the local feature and the global feature to obtain an output of the first submodule.15.The method of claim 14, wherein the first feature representation and the second feature representation are obtained through a single convolution layer and channel wise separation, orthe first feature representation and the second feature representation is obtained through two individual convolution layers.16.The method of any of claims 14-15, wherein the local feature and the global feature is combined through channel wise concatenation and a convolution layer.17.The method of any of claims 14-16, wherein the first submodule comprises at least one of the following: one or more additional layers, or one or more residual blocks.18.The method of any of claims 11-17, wherein the first neural network further comprises a second normalization layer and a second submodule, and an output of the first submodule is processed through the second normalization layer and the second submodule.19.The method of claim 18, wherein the second normalization layer comprises one of the following: a layer normalization layer, a batch normalization layer, an instance normalization layer, or a group normalization layer.20.The method of any of claims 18-19, wherein the second submodule comprises a stack of a convolution block, an activation function, a partition operation, a fast Fourier transform (FFT) block, a weight factor multiplication operation, an inverse fast Fourier transform (IFFT) block and a merging operation.21.The method of claim 20, wherein the convolution block comprises a convolution layer or depth wise convolution.22.The method of any of claims 20-21, wherein the activation function comprises one of the following: a Gaussian error linear unit (GELU) , a rectified linear unit (ReLU) , a leaky rectified linear unit (LeakyReLU) , an exponential linear unit (ELU) , a scaled exponential linear unit (SELU) .23.The method of any of claims 20-22, wherein an output of the second submodule is generated by:processing a feature representation inputted to the second submodule with the convolution block and the activation function;dividing the processed feature representation into a first channel feature part and a second channel feature part;generating a first modulated feature and a second modulated feature based on the first channel feature part and the second channel feature part; andgenerating the output of the second submodule based on the first modulated feature and the second modulated feature.24.The method of claim 23, wherein generating the first modulated feature and the second modulated feature comprises:generating a first gating weight based on the first channel feature part;generating a second gating weight based on the second channel feature part;generating the first modulated feature by applying the second gating weight to the first channel feature part; andgenerating the second modulated feature by applying the first gating weight to the second channel feature part.25.The method of claim 24, wherein the first gating weight is generated by processing the first channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function, andthe second gating weight is generated by processing the second channel feature part with a stack of a linear block, a nonlinear activation function, a linear block, and a nonlinear gating function.26.The method of claim 25, wherein the nonlinear activation function comprises a GELU, and the nonlinear gating function comprises a Sigmoid function.27.The method of any of claims 23-26, wherein generating the output of the second submodule comprises:combining the first modulated feature and the second modulated feature to obtain a fused feature;applying a multi-layer perceptron (MLP) with an activation function to the fused feature to obtain a refined feature; andgenerating the output of the second submodule based on the refined fused feature.28.The method of claim 27, wherein the refined fused feature is projected back to an original channel dimension to obtain the output of the second submodule.29.The method of any of claims 11-28, wherein each of the at least one NN-based module further comprises a second neural network configured to perform an inverse process of the first neural network.30.The method of any of claims 18-29, wherein each of the at least one NN-based module comprises more than one first submodule and more than one second submodule.31.The method of any of claims 1-30, wherein during a training process of the NN-based model, the NN-based model is trained in a distributed manner at a plurality of devices, each of the plurality of devices is assigned to a distinct bitrate and determines respective gradients, the gradients of the plurality of devices are synchronized and aggregated for updating parameters of the NN-based model.32.The method of claim 31, where at each of the plurality of devices, a loss function for training the NN-based model is as follows: L=R+λ×Dwherein L represents to a training loss, R represents to bitrate, D represents a distortion, and λ represents a weighting factor.33.The method of claim 32, wherein values of the weighting factor λ are different for different training examples.34.The method of any of claims 32-33, wherein values of the weighting factor λ are different for different samples on a batch size dimension.35.The method of any of claims 32-34, wherein a value of the weighting factor λ is determined from a set of candidate values.36.The method of claim 35, wherein the set of candidate values comprises 85, 170, 380, and 840, orwherein the set of candidate values comprises all integers within a value range from a first integer to a second integer.37.The method of claim 36, wherein the first integer is 85, and the second integer is 840.38.The method of claim 32, wherein values of the weighting factor λ are different for different devices of the plurality of devices.39.The method of any of claims 31-38, wherein the training process is applied at a finetuning stage of the NN-based model, orthe training process is applied from the beginning of training the NN-based model.40.The method of any of claims 1-39, wherein the visual data comprise a video, a picture of the video, or an image.41.The method of any of claims 1-40, wherein the conversion comprises encoding the visual data into the bitstream.42.The method of any of claims 1-40, wherein the conversion comprises decoding the visual data from the bitstream.43.The method of any of claims 1-40, wherein the conversion comprises: generating the bitstream from the visual data, andthe method further comprises: storing the bitstream in a non-transitory computer-readable recording medium.44.An apparatus for visual data processing comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform a method in accordance with any of claims 1-43.45.A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of claims 1-43.46.A non-transitory computer-readable recording medium storing a bitstream of visual data which is generated by a method performed by an apparatus for visual data processing, wherein the method comprises:generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data.47.A method for storing a bitstream of visual data, comprising:generating the bitstream of the visual data with a neural network (NN) -based model, wherein the NN-based model comprises at least one NN-based module, each of the at least one NN-based module is configured to process an input feature representation of the NN-based module based on a local feature and a global feature associated with the input feature representation, and the input feature representation comprises a feature representation associated with the visual data; andstoring the bitstream in a non-transitory computer-readable recording medium.48.A computer program for performing a method in accordance with any of claims 1-43 when the computer program runs on a computer.49.A computer program product tangibly stored in a computer storage medium and comprising computer-executable instructions that, when executed by a device, cause the device to perform a method in accordance with any of claims 1-43.