Method, apparatus, and medium for visual data transcoding

Abstract

Embodiments of the present disclosure provide a solution for visual data transcoding. A method for visual data transcoding is proposed. The method includes: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream 5 of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

Description

METHOD, APPARATUS, AND MEDIUM FOR VISUAL DATA TRANSCODING FIELDS

[0001] Embodiments of the present disclosure relate generally to visual data transcoding techniques, and more particularly, to neural network-based visual data transcoding.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 gained 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 being improved, neural network-based video compression has become an actively developing research area. However, neural network-based image / video transcoding is generally expected to be researched.SUMMARY

[0003] Embodiments of the present disclosure provide a solution for visual data transcoding.

[0004] In a first aspect, a method for visual data transcoding is proposed. The method comprises: reconstructing visual data from a first bitstream of the visual data: processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN -based coding scheme.

[0005] Based on the method in accordance with the first aspect of the present disclosure, during the transcoding process involving a bitstream generated according to an NN-based coding scheme, after the visual data is reconstructed from the first bitstream, the reconstructed visual data is processed with the NN-based module, and then the processed visual data is encoded into the second bitstream. Compared with the conventional solution where the reconstructed visual data is directly encoded into a further bitstream, the proposed method can advantageously utilize the processing with NN-based module to optimize the reconstructed visual data during the transcoding process, and thus bridging the gap between the two coding schemes involved in the transcoding process. Thereby, the transcoding quality can be improved.

[0006] In a second aspect, an apparatus for visual data transcoding is proposed. The apparatus comprises a processor and a non -transit ory 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.1 F1257165PCT

[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 visual data which is generated by a method performed by an apparatus for visual data transcoding The method comprises: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

[0009] In a fifth aspect, a method for storing a bitstream of visual data is proposed. The method comprises: reconstructing visual data from a first bitstream of the visual data: processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data: and encoding the processed visual data inlo a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

[0010] 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 matterBRIEF DESCRIPTION OF THE DRAWINGS

[0011] Through the following detailed description with reference to the accompanying drawings, the above and other objectives, features, and advantages of example embodiments of the present disclosure will become more apparent In the example embodiments of the present disclosure, the same reference numerals usually refer to the same components

[0012] Fig. 1A illustrates a block diagram that illustrates an example visual data coding system, in accordance with some embodiments of the present disclosure

[0013] Fig. 1B is a schematic diagram illustrating an example transform coding scheme.

[0014] Fig. 2 illustrates example latent representations of an image.

[0015] Fig. 3 is a schematic diagram illustrating an example autoencoder implementing a hyperprior model.

[0016] 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.

[0017] Fig. 5 illustrates an example encoding process.

[0018] Fig. 6 illustrates an example decoding process.

[0019] Fig. 7 illustrates a learning-based image codec architecture

[0020] Fig. 8 illustrates an example Joint Photographic Experts Group (JPEG) Artificial Intelligence (Al) decoder structure.

[0021] Fig. 9 illustrates an example JPEG AI decoder for one component.

[0022] Fig. 10 illustrates an example of transcoding for different resolutions, bitrates, or compatible codecs available on playback devices.

[0023] Fig. 11 illustrates an example scenario where the JPEG-coded files cannot to be decoded on devices with JPEG Al image codec.2 F1257165PCT

[0024] Fig. 12 illustrates an example transcoding scheme for transcoding JPEG coded files to JPEG Al decodable files, in accordance with some embodiments of the present disclosure

[0025] Fig 13 illustrates an example transcoding scheme where filtering is used prior to the second decoder, in accordance with some embodiments of the present disclosure[0026JFig. 14 illustrates an example transcoding scheme where at least two encoders of the second codec are used, in accordance with some embodiments of the present disclosure, wherein one encoder is the original one, and the others are retrained for transcoding purpose.00271Fig. 15 illustrates an example of the transcoding module inserted between the first codec and the second codec, in accordance with some embodiments of the present disclosure

[0028] Fig. 16A illustrates an example architecture of the transcoding module, in accordance with some embodiments of the present disclosure.

[0029] Figs 16B and 16C illustrate an example of the local-global modulation module (LGMM) and the dynamic fusion module (DFM) of the transcoding module as shown in Fig. 16A, respectively.

[0030] Fig. 17 illustrates a flowchart of a method for visual data transcoding in accordance with some embodiments of the present disclosure.

[0031] Fig. 18 illustrates a flowchart of a method for processing the reconstructed visual data in accordance with some embodiments of the present disclosure.

[0032] Fig. 19 illustrates a flowchart of a method for visual data transcoding in accordance with embodiments of the present disclosure.

[0033] Fig 20 illustrates a block diagram of a computing device in which various embodiments of the present disclosure can be implemented.

[0034] Throughout the drawings, the same or similar reference numerals usually refer to the same or similar elementsDETAILED DESCRIPTION

[0035] Principle of the present disclosure will now be described with reference to some embodiments. It is to be understood that these embodiments 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.[0036 J 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.

[0037] References in the present disclosure to “one embodiment.” “an embodiment,” “an example embodiment,” and the like indicate that the embodiment described may include a particular feature, structure, or characteristic, but it is not necessary that every embodiment includes the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an example embodiment, 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 embodiments whether or not explicitly described.3 F1257165PCT

[0038] 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 embodiments. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.

[0039] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. 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 addi tion of one or more other features, elements, components and / or combinations thereof.Example Environment

[0040] 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 visual data encoding device, and the destination device 120 can be also referred to as 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 visual data source 112, visual data encoder 114, and an input / output (I / O) interface 116.

[0041] The visual data source 112 may include a source such as visual data capture device Examples of the visual data capture device include, but are not limited to, an interface to receive visual data from visual data provider, a computer graphics system for generating visual data, and / or a combination thereof.

[0042] The visual data may comprise one or more pictures of a video or one or more images. 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 visual data. The coded picture is a coded representation of a picture. The associated visual 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.

[0043] The destination device 120 may include an I / O interface 126, 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 4 F1257165PCT120, or may be external to the destination device 120 which is configured to interface with an external display device.

[0044] The visual data encoder 114 and the visual data decoder 124 may operate according to visual data coding standard, such as video coding standard or still picture coding standard and other current and / or further standards

[0045] Some exemplary embodiments of the present disclosure will be described in detailed hereinafter. It should be understood that section headings are used in the present document to facilitate ease of understanding and do not limit the embodiments disclosed i n a section to only that section. Furthermore, while certain embodiments are described with reference to Versatile Video Coding or other specific visual data codecs, the disclosed techniques are applicable to other coding technologies also. Furthermore, while some embodiments describe coding steps in detail, it will be understood that corresponding steps decoding that undo the coding will be implemented by a decoder. Furthermore, the term visual data processing encompasses visual data coding or compression, visual data decoding or decompression and visual data transcoding in which visual data are represented from one compressed format into another compressed format or at a different compressed bitrate.1. Brief SummaryA neural network -based image and video compression transcoding method is described in the present disclosure. The image or video codec, namely the target codec, comprises multiple neural networks, wherein the encoder encodes the original image / video into bitstreams, and the decoder reconstructs the image / video. However, it is impossible for the codec to decode the bitstream file encoded with another codec. The present disclosure herein proposes a transcoding scheme to support the functionality of decoding with the second codec, when the bitstreams are generated with the first codec.2. BackgroundThe 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 gained 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 being improved, 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 problem2.1. Image / Video compressionImage / 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 losslessly reconstructing the original image / video, termed lossless compression and lossy compression Most of the efforts are devoted to 5 F1257165PCTlossy 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 (V CEG) 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 HE VC.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 networksNeural networks, also known as artificial neural networks (ANN), are the computational models used in machine learning technology which tire 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 compressionF1257165PCTExisting neural networks for image compression methods can be classified 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 ModelingAccording 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 dimensionalityFollowing the predictive coding strategy, one way to model p(x) is to predict pixel probabilities one by one m a raster scan order based on previous observations, where x is an image.p(x) = p(x,)p(x2|1)...p(Xj|X1,..., Xi_1)...p(xmx„|X1,..., Xmxn-1) 0) 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|x!-fc,...,xi_...p(xmxxmxn-fc,...,xmxn-1) (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(xt) given its context xvx2,....x^. 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 is a feed-forward network with a single hidden layer. A similar work is presented in [8], where 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(x 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 using 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 is proven to be good at modeling sequential data. The spatial variant of LSTM is used for 7 F1257165PCTimages 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: they consider pixels as discrete values (e.g., 0, 1,..., 255) and predict a multinomial distribution over the discrete values: they deal with color images in RGB color space; they work well 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- ay 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. PixelSNAlL 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:p(x\h) = np^pCNki, (3) where h is the additional condition and p(x) = p( / i)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-encoderAuto-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. IB is a schematic diagram illustrating an example transform coding scheme. Fig. IB is an illustration of a typical transform coding scheme. The original image x is transformed by the analysis network gato 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 y is then inversely transformed by a synthesis network gsto obtain the reconstructed image x. The distortion is calculated in a perceptual space by transforming x and x with the function gp.It is intuitive to apply’ auto-encoder network to lossy image compression It is only need 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 m backpropagation while training the neural networks. Second, the objective under compression scenario is different since both the distortion and the rate need to be take into consideration. Estimating the rate 8 F1257165PCTis 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 areaThe prototype auto-encoder for image compression is in Fig IB, which can be regarded as a transform coding strategy The original image x is transformed with the analysis network y — S'a( ), where y is the latent representation which will be quantized and coded. The synthesis network will inversely transform the quantized latent representation y back to obtain the reconstructed image x gsy. The framework is trained with the rate-distortion loss function, i.e, £ = D +. R. where D is the distortion between x and x, R is the rate calculated or estimated from the quantized representation y, and A 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 functionIn 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. They use binary quantization 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-weigbted loss function is also designed, and spatially adaptive bitrates mechanism is enabled. They achieve better results than BPG on Kodak image dataset using MS-SSIM as evaluation metric.A general framework was designed for rate-distortion optimized image compression. They use multiary quantization 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. They add random uniform noise to stimulate the quantization during training and use the differential entropy of the noisy codes as a proxy for the rate. They use generalized divisive normalization (GDN) 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 they use 3 convolutional layers each followed by a down-sampling layer and a GDN layer as the forward transform. Accordingly, they use 3 layers of inverse GDN 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. They transform the latent representation y with a subnet hato z -- h.ay) and z will be quantized and transmitted as side information. Accordingly, the inverse transform is implemented with a subnet / tsattempting to decode from the quantized side information z to the standard deviation of the quantized y, which will be further used during the arithmetic coding of y. 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 9 F1257165PCTset using PSNR as evaluation metric.2.3.3. Hyper Prior Mode!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 ga(x, 0g) into a latent representation y, which is then quantized to form y. Because y is discrete-valued, it can be losslessly compressed using entropy coding techniques such as arithmetic coding and transmitted as a sequence of bits.Fig. 2 illustrates example latent representations of an image. As evident from the middle left and middle right image of Fig 2, there are significant spatial dependencies among the elements of y Notably, their scales (middle right image) appear to be coupled spatially An additional set of random variables z are introduced to capture the spatial dependencies and to further reduce the redundancies Tn this case the image compression network is depicted in Fig 3.Fig. 3 is a schematic diagram illustrating an example autoencoder implementing a hyperprior model. In Fig 3, the left hand of the models is the encoder gaand decoder gs(explained in section 2 3.2). The right-hand side is the additional hyper encoder haand hyper decoder hsnetworks that are used to obtain z. 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 h.a, summarizing the distribution of standard deviations in z. z is then quantized (z), compressed, and transmitted as side information. The encoder then uses the quantized vector z to estimate ff, the spatial distribution of standard deviations, and uses it to compress and transmit the quantized image representation y. The decoder firs! recovers z from the compressed signal. It then uses hsto obtain o, which provides it with the correct probability estimates to successfully recovery as well. It then feeds j? intoto 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 y 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.Fig. 3 shows a network architecture of an autoencoder implementing the hyperprior model. The left side shows an image autoencoder network, the right side corresponds to the byperprior subnetwork. The analysis and synthesis transforms are denoted as gaand 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 (z) which comprises information about the probability? distribution of the samples of the quantized latent y. z is included in the bitstream and transmitted to the receiver (decoder) along with y.2.3.4. Context ModelAlthough the hyperprior model improves the modelling of the probability? distribution of the quantized latent y, 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 10 F1257165PCTmodel along with a hyperprior and the autoencoder. The following Table illustrates meaning of different symbolsTable - Illustration of symbolsComponent SymbolInput Image XEncoderLatents yLatents (quantized) yDecoder g(y. &d)Flyper Encoder fh(y. ^he)Hypcr-Latcnts zI Iv per-L atent s (quantized) zHyper Decoder 9h ®hd)Context Model 9 cm (y ® cm )Entropy Parameters 9 ep I* ' ®ep)Reconstruction XA 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 y, 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 fi and scale (or variance) c 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 aritlimetic encoder (AE) module. In the decoder the gaussian probability model is utilized to obtain the quantized latents y from the bitstream by arithmetic decoder (AD) module.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. For example. Fig. 4 shows 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 (y) and quantized byper-latents (z), 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 asand er).2.3.5. The encoding process using joint auto-regressive hyper prior modelFig. 5 illustrates an example encoding process. For example, Fig 5 shows to the state-of-the-art compression 11 F1257165PCTmethod. Tn this section and the next, the encoding and decoding processes will be described separately.The figure above 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 obtain the quantized latent (y ) y is then converted to a bitstream (bitsl) using an arithmetic encoding module (denoted AE). The arithmetic encoding block converts each sample of the y into a bitstream (bitsl) one by one, in a sequential orderThe modules hyper encoder, context, hyper decoder, and entropy parameters subnetworks are used to estimate the probability distributions of the samples of the quantized latent y The latent y is input to hyper encoder, which outputs the hyper latent (denoted by z) The hyper latent is then quantized (z) 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 y. The information that is generated by the Entropy? Parameters typically include a mean / i and scale (or variance) c parameters, that are together used to obtain a gaussian probability distribution A gaussian distribution of a random variable x is defined as f(x) =— cry27r wherein the parameter fl is the mean or expectation of the distribution (and also its median and mode), while the parameter or 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 ( E) module. The quantized latent y is typically a matrix composed of many samples. The samples can be indicated using indices, such as (i,j,k| or [i,j ] depending on the dimensions of the matrix y The samples y[i,j] 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;\E to encode the samples into bitstream), the context module generates the information pertaining to a sample y [i,j], 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, y into bitstream (bitsl).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 modelFig 6 illustrates an example decoding process For example. Fig 6 depicts the state-of-the-art decoding process In the decoding process, the decoder first receives the first bitstream (bits 1 ) and the second bitstream (bits2) that 12 F1257165PCTare 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 z, 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 z that was generated by the encoder can be reconstructed at the decoder without any change.After obtaining of z, 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 m encoder), which is essential for reconstructing the quantized latent y without any loss As a result, the identical version of the quantized latent y 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 bitsl. 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 y 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. Neural netw orks for video compressionSimilar 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 is 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.2.4.1. Low -latencyThe 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- F1257165PCTencoders 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 firstly 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 performs 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 HE VC 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 belter performance than H.264.A multi-resolution representation for optical flows is proposed. Concretely, the motion estimation network produces multiple optical Hows with different resolutions and let the network to learn which one to choose under the loss function. The performance is better than H.265.2.4.2. Random accessA 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. They perform motion compensation 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 14 F1257165PCTdeterministic 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. Preliminaries2.5.1. Colors separation and conditional codingIn one example the primary and secondary color components of an image are coded separately, using networks with similar architecture, but different number of channels as shown in 8. All boxes with same names are sub¬ networks with the similar architecture, only input-output tensor size and number of channels are different. Number of channels for primary component is Cp= 128, for secondary7components is Cs-- 64. The vertical arrows (with arrowhead pointing downwards) indicate data flow related to secondary color components coding. Vertical arrows show data exchange between primary and secondary components pipelines.The input signal to be encoded is notated as x, latent space tensor in bottleneck of variational auto-encoder is y. Subsenpt “ Y” indicates primary component, subsenpt “U V” is used for concatenated secondary components, there are chroma components.First the input image that has RGB color format is converted to primary (Y) and secondary components (UV). The primaiy componentis coded independently from secondary components xllvand the coded picture size is equal to input / decoded picture size. The secondary components are coded conditionally, using xYas auxiliary information from primary' component for encoding xuvand using yYas a latent tensor with auxiliary information from primary component for decoding yLIVreconstruction. The codec structure for primary component and secondary components are almost identical except the number of channels, size of the channels and the several entropy models for transforming latent tensor to bitstream, therefore primaiy7and secondary latent tensor will generate two different bitstream based on two different entropy models. Fig. 7 illustrates a learning-based image codec architecture. Prior to the encoding xY, xuvgoes through a module which adjusts the sample location by down-sampling (marked as “sV” on Fig. 7), this essentially means that coded picture size for secondary component is different from the coded picture size for primary component. The scaling factor s is variable, but the default scaling factor is s = 2. The size of auxiliary input tensor in conditional coding is adjusted in order the encoder receives primaiy and secondary components tensor with the same picture size. After reconstruction, the secondary7component is rescaled to the original picture size with a neural -network based upsampling filter module (“NN-color filter st” on Fig. 7), which outputs secondary components up-sampled with factor sFig. 8 illustrates an example Joint Photographic Experts Group (JPEG). Artificial Intelligence (Al) decoder structure. The example in Fig. 8 exemplifies an image coding system, where the input image is first transformed into primary (Y) and secondary components (UV) The outputs, xuvare the reconstructed outputs corresponding to the primary7and secondary7components. At the end of the processing, xY, xuvare converted back to RGB color format. Typically, the xuvis downsampled (resized) before processing with the encoding and decoding modules (neural networks). For example, the size of the xuymight be reduced by a factor of 50% in each of the vertical and horizontal dimensions. Therefore, the processing of the secondary component includes approximately 50% x 50% = 25% less samples, therefore it is computationally less complex15 F1257165PCT2.5.2. The JPEG Al image coding standardAt the time of writing, the JPEG Al image coding standard is an image coding standard that is being standardized by the JPEG Working Group (WG), which is WG 1 of ISO / IEC JTC 1 SC 29 The ISO / IEC number for the. JPEG Al standard is ISO / IEC 6048 The latest JPEG AT draft specification is included in. JPEG output document WG1N100602.The design in the latest JPEG Al draft specification utilizes some NN-based image coding methods described mentioned above. Some of the features in the latest JPEG Al specification are described or summarized below.2.5.2.1. Functional overview of the JPEG Al decoding processThe overall decoder architecture in shown in Fig. 8 Data (tensors and streams) are shown inside the "white” boxes, neural network modules necessary for decoding are shown in grey shadowed boxes, switchable tools are shown in purple shadowed boxes.For primary and secondary colour components code streams can be parsed independently and reconstructed using modules consisting of same sequence of same neural -network layers, with the only difference in sizes on input tensors and number of tensor channels. Single component decoder is shown in Fig. 9. Fig. 9 illustrates am example JPEG Al decoder for one componentFirst stream z shall be parsed by loss-less entropy decoder (me — CANS decoder). The probability distribution for loss-less coding of z is assumed to be Gaussian with pre-trained parameters (part of the trained model), Commulative Distribution Function (denoted on Fig. 9 as CDF(z)) computed based on those pre-trained parameters is used in loss-less entropy decoder.Decoded hyper-prior tensors z is used as an input for two different processes: Hyper Decoder and Hyper Scale Decoder.Then stream y shall be parsed by loss-less decoder (me — tANS decoder). The probability distribution for parsing f is assumed to be Gaussian with zero mean value and standard deviation given as an output if following steps: Hyper Scale Decoder (section 10.3) outputs tensors of standard deviation in log-domain n / ^[C, h4, w4], then it is scalled according to the rate control parameter p inside Sigma Scale to produce as / 'ff, and then masked and scalled according to RVS parameters section inside Adaptive Sigma Scale producing I" Finally, tensor I"ffvalues are quantized (converted to the index of probility distribution table). According to the rules, specified by SKIP Mode some elements of residual tensor are skipped (not encoded / decoded) and replaced by zeros in Decoder SKIP module, which receives parsed set of syntax elements {,v} from tANS Decoder, mask_sigma from SKIP Mask generation module and outputs re-shaped to 3D shape reconstructed residual tensor r [C, h4, w4]. At decoder side the residual r is scaled by Inverse Gam Unit according to the parameter / ?, producing r'. Then residual tensor is scaled in invRVS (Inverse Residual and Variance Scale) moduleforming residual tensor r" This is used for reconstructed latent tensor yHyper decoder generates explicit -prediction input to Multi-stage Context Model -MCM, which is eight stages neural network process, which also takes reconstructed residual r" as an input and outputs latent space tensors y'. After Patent Scaling Before Synthesis-LSBS reconstructed latent space tensor y is ready for signal reconstruction. Latent tensors reconstructions for primary and secondary components are independent from each other.Reconstructed latent space tensors j>[C, / i4,w4] is an input of Synthesis Transform. Another input of Synthesis 16 F1257165PCTtransform is auxiliary tensor y[, h4, w4]. For secondary component Synthesis the auxiliary tensor is generated from primary component recon stnicted latent tensor For primary component no auxiliary tensor is used Depending on input picture height H and width IV and scaling factors for primary (sy) and secondary (suv) components sizes of tensors are shown in Table 1. For primary' component the parameter Cd— 0 This means that primary' component’s Synthesis transform receives no auxiliary information (reconstructed independently) For secondary' component Cd= 160, the auxiliary' for secondary transform synthesis is, which is re-sampled by integer factor suv / sY(using nearest neighbour down-sampling or nearest neighbour up-sampling) reconstructed latent space tensor of primary' component yYTable 1 Tensor size parameters for primaiy and secondary' components decodingPrimaiy “Y” component Secondaty “UV” component HinH ceil(H / sver)W ceil(W / shor)2hd, d = 0,...,6 cet7(Hin / 2d) ceil(Hin / 2f'd<‘i^d:d~1) wd, d 0,...,6 ceil(Wir, / 2d) ceil(Win / 2(d<^?d'd~1)C 160 32 c396 + 32 * opldx 64 C2= Q 64 + 64 * opldx 64 Q 0 160Synthesis transform for primary and secondary component consists of same neural network layers, the only difference is the size of input tensor and number of tensor channels. Synthesis transform outputs tensor x [Cin, Hin, Win] (tensor sizes are listed in Table 1 ).As shown on Fig. 8 after synthesis transform, the primary and secondary components go to the Enhancement filters and output format conversion processing module, which includes re-sampling, inverse color conversion and set of filters.2.5.3. TranseodingImage / video transcoding is the process of converting a video / image file, i.e., the bitstream file, from one format to another. It involves changing an image / video file’s encoding format, resolution, bitrate, or other parameters to ensure compatibility with different devices, platforms, or internet connection capabilities. Transcoding is key in video production and distribution. Fig. 10 illustrates an example of transcoding for different resolutions, bitrates, or compatible codecs available on playback devices. As the transcoding example shown in Fig. 10, the image is encoded with the first codec encoder followed by decoding with the first codec decoder, which is part of the transcoder. The reconstructed image is re-encoded with the second codec encoder, and the resulting bitstream files are distributing over the internet. When the users request to display, the second codec, decoder is used to reconstruct the imageFig. 10 show's an example of transcoding for different resolutions, bitrates, or compatible codecs available on playback devices17 F1257165PCTImage / video transcoding technologies develop with the user content digitalization process when multiple generations of imageA’ideo codecs have been developed to accommodate the increasing visual content. Whenever a new codec is released, transcoding technique needs to be applied to convert the bitstreams / files coded with another codec, e.g.. the previous generation codec, to the decodable format for the latest codec There are primarily the following reasons why image / video transcoding is important.1 ) Compatibility. Different devices and platforms have support for different resolutions, bitrates, codecs, and container formats Transcoding enables seamless conversion into desired formats that are compatible with a wider range of devices and platforms.2) Optimization. Transcoding allows videos to be optimized for different devices (desktops, smartphones, tablets, or smart TVs) which all have different definitions for 'optimal’ settings. By transcoding videos into resolutions and bitrates that ensure the smoothest playback for each at the best possible image quality, you ensure a better viewing experience for users.3) Streaming. Transcoding is business critical for video streaming services like YouTube, Netflix, Amazon Prime Video. These seivic.es need to support a wide range of devices, operating systems, browsers. They use transcoding as a solution - optimizing their videos into multiple formats and resolutions to ensure smooth playback on a wide range of user platforms and bandwidths.4) File size reduction. Transcoding videos can also reduce the file size of the video, making it easier and quicker to upload, share, or stream. This is especially important when dealing with large videos. 5) Accessibility. Closed captions and subtitles can be added during the transcoding process, enabling users with hearing impai rments to understand the content of the video. Videos can also be transcoded with audio streams in different languages, to cater to non-native speakers.3. Problems3.1, The core problemAt the time of this draft, JPFG Al image coding standard is a neural network -based image coding standard under development Prior to this standard, there have been many other image coding standards, such as JPEG, JPEG 2000. As exemplified in Fig 11, the problems arise when the terminal devices only have JPEG A. I codec while the image files are coded with other codecs, for example JPEG. Although JPEG Al has been significantly efficient compared to the predecessors (. JPEG,. JPEG 2000), without transcoder making the JPEG Al coding standard difficult for wide deployment.Fig. 11 shows an example that the JPEG-coded files cannot to be decoded on devices with JPEG Al image codec. In addition to the compatible codec use case, there are other scenarios that the transcoding is a necessity.* Changing the image resolution or bitrate.« Enhance or optimize the reconstructed image quality depending on the bitrate or application types. « Reduce the bitstream file size on the storage devices.® Adjust the bitrate depending on the internet bandwidth.4. EmbodimentsThe detailed embodiments below should be considered as examples to explain general concepts. These embodiments should not be interpreted in a narrow way. Furthermore, these embodiments can be combined m any manner.18 F1257165PCTFig. 12 illustrates an example to transcode JPEG coded files to JPEG Al decodable files. An example is shown in Fig. 12 wherein a transcoder is used to convert the bitstream files generated with the first codec to the compatible formats such that the second codec can decode to obtain the correct reconstructed images4.1. Core of the present disclosureThe core of the present disclosure is to offer the transcoding solution for a neural network -based image codec, such as JPEG Al image coding standard, wherein the coded files using the first codec can be correctly decoded with the second codec decoder to obtain the correct reconstructed images For example. Fig. 13 shows an example showing filtering is used prior to the second codec decoder5. Embodiments1. According to the invention, a transcoding solution is proposed to convert ti first bitstream files generated by a first codec to a second bitstream with the compatible format for a second codec decoder so that the second decoder can reconstruct the images.a. In one example, the second codec may be neural network -based codec, such as JPEG Al image coding standard.i. Alternatively, the second codec may be traditional / non-neural-net work-based (NN- based) codec, such as. JPEG, JPEG 2000, H.264, H.265 or H.266.b. In one example, the first codec may be traditional / non-neural-network -based codec, such as JPEG, JPEG 2000, H.264, H.265 or H.266.i. Alternatively, the first codec may be neural network -based codec, such as JPEG Al image coding standard.c. Tn one example, the second codec may remain unchanged, and an intermediate-processing module is used between the first codec decoder and the second codec encoder.i In one example, the second codec is a NN-based coded and the NN weights are not ch ngedii. In one example as shown in Fig. 13, the intermediate processing may be filtering.1. In one example, the filtering may be NN-basedd. In one example as shown in Fig. 14, the NN-based second codec may comprise multiple sets of weights. Fig. 14 illustrates an example transcoding scheme where at least two encoders of the second codec are used, in accordance with some embodiments of the present disclosure, wherein one encoder is the original one, and the others are retrained for transcoding purpose.i. For example, a first set of NN weights may be for purpose other than transcoding. ii. For example, a second set of NN weights may be for purpose of transcoding.1. In one example, the NN wei hts may depend on the first codec.2. An indication may be signaled in the bitstream to indicate which set of NN weights is used.e. As shown in Fig. 14, the multiple sets of NN weights of the second codec may be organized with different architectures and / or number of layers and / or types of neural network layers. 2. According to the invention a transcoder is used to assist the decoder to reconstruct an image, wherein the transcoder and decoder both consist of neural network layers.F1257165PCTa. An indication might be included in the bitstream to indicate if the transcoder is used in processing or if it is skipped.b Multiple encoders of the second codec may be used and an indication might be included in the bitstream to indicate which one to be used in the decoding processAccording to the invention a transcoding module may be inserted between the first codec and the second codec. The transcoding module accepts the decoded image as the input and outputs the processed image, which is fed into the encoder of the second codec.a. in one example, the first codec may be traditional image coding codec, such as JPEG, JPEG 2000, H.264 / AVC intra coding, H.265 / HEVC intra coding, H.266 / V VC intra coding, etc, transcoding module is neural network-based module, and the second coded may be neural network-based image codec, such as JPEG Al.b. In one example, as shown in Fig. 15, the transcoding module may be designed to explore both the local and global features from the decoded image of the first codec. Fig. 15 illustrates an example of the transcoding module (e.g., an Efficient Local-Global Collaboration-Driven Transcoding Network (ELGC-TransNet)) inserted between the first codec and the second codec, in accordance with some embodiments of the present disclosure.c. in Fig. 15, JPEG, IIEIF are used as examples of the first codec, however, the first codec may be many other codecs. JPEG Al is used as an example of the second codec.d. In one example, the detailed design of the transcoding module ELGC-TransNet may be designed as shown in Figs 16A-16C. The transcoding module consists of two branches composed by stacked local-global modulation modules (LGMM) and dynamic fusion modules (DFM) to explore the global as well as local information and fuse them together. Fig. 16A illustrates an example of the detailed architecture of the transcoding module, in accordance with some embodiments of the present disclosure Figs 16F> and 16C illustrate an example of the local-global modulation module (LGMM) and the dynamic fusion module (DFM) of the transcoding module as shown in Fig. 16A, respectively.e. In Figs. 16A-16C, the designs of these two branches, i.e., stacked DFM and stacked LGMM, may vary as long as they are able to explore both local and global information and fuse the information together. Figs. 16A-16C illustrates the detailed architecture of the ELGC-TransNet. There are two branches composed by slacked local-global modulation module (LGMM) and the dynamic fusion module (DFM), which are used to extract and fuse both global and local information and fuse the inherent rel tionships.In one example, at least one indication may be signaled in a second bitstream to indicate the type of the original picture from which it is encoded.a. The type may be a raw picture.b. The type may be a picture decoded from a first bitstream coded by a specific codec, such as JPEG, JPEG 2000, H.264, H.265 or H.266.i. The kind of specific codec may be signaled in the second bitstream, represented by a reference value such as 0, or 1, etc.20 F1257165PCTc. A post processing may be applied on the picture decoded from the second bitstream.i. In one example, whether to and / or how to apply the post processing may depend on the type of the original picture from which it is encoded1 For example, a specific post processing may be applied if the original picture is decoded from a first bitstream coded by a specific codec.li. In one example, the post processing may be filtering.iii. In one example, the post processing may be NN-based.5. According to the invention an indicator might be included in the bitstream to indicate:a. Selection of model parameters (e.g. weights of convolution layers) to be used in decoding of a transcoded image.i. For example, the indicator might indicate a selection of model coefficients (e.g. parameters or weights') of a processing layer.b. Selection of an entropy coding mode.i. For example, the indicator might indicate a selection of a table to be used in entropy decodingi i. In another example the indicator might indicate a selection an initialization method of an entropy decoding.c. Selection of a synthesis transform.i. For example, if the indicator indicates that the bitstream is obtained by transcoding, a different synthesis transform might be appliedd. Selection of a color transform.i. The indicator might indicate selection of a color transform or an inverse colour transform.General aspects6. Whether to and / or how to apply the disclosed methods above may be signalled at block level / sequence level / group of pictures level / picture level / slice level / tile group level, such as in coding structures of CTU / CU / TUZPU / CTB / CB / TB / PB, or sequence header / picture header / SPS / VPS / DPS / DCI / PPS / APS / slice header / tile group header.7. Whether to and / or how to apply the disclosed methods above may be dependent on coded information, such as block size, colour format, single / dual tree partitioning, colour component, slice / picture type.8. The proposed methods disclosed in this document may be used in other coding tools which require chroma fusion.9. A syntax element disclosed above may be binarized as a flag, a fixed length code, an EG(x) code, a unary code, a truncated unary code, a truncated binary code, etc. It can be signed or unsigned.10. A syntax element disclosed above may be coded with at least one context model. Or it may be bypass coded.11. A syntax element disclosed above may be signaled in a conditional way.a. The SE is signaled only if the corresponding function is applicable.b. The SE is signaled only if the dimensions (width and / or height) of the block satisfy a condition.F1257165PCT12. A syntax element disclosed above may be signaled at block level / sequence level / group of pictures level / picture level / slice level / tile group level, such as in coding structures of CTU / CU / TU / PU / CTB / CB / TB / PB, or sequence header / picture header / SPS / VPS / DPS / DCI / PPS / APS / slice header / tile group header

[0046] More details of the embodiments 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.

[0047] Existing transcoding schemes have focused primarily on the compatibility between traditional non-NN-based image / video compression standards through color space conversion, resolution scaling, quantization adjustments, post-processing enhancement and / or the like They are not designed for transcoding scenarios involving NN-based coding scheme, and may lead to inefficient bit allocation and decreased semantic accuracy.

[0048] To solve the above problems and some other problems not mentioned, visual data transcoding solutions as described below are disclosed. The embodiments 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 embodiments can be applied individually or combined in any manner

[0049] Fig. 17 illustrates a flowchart of a method 1700 for visual data transcoding in accordance with embodiments of the present disclosure. The method 1700 starts at 1702, where visual data is reconstructed from a first bitstream of the visual data By way of example rather than limitation, the reconstruction may be performed with a decoder corresponding the above-mentioned first codec.

[0050] At 1704, the reconstructed visual data is processed with a neural network (NN)-based module to obtain processed visual data The above-mentioned transcoding module may be an example implementation of the NN-based module As used herein, an NN-based module may be a module based on neural network technologies For example, an NN-based module may specify sequence of neural network modules (also called architecture) and module 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 module described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any way. In addition, the function of the NN-based module may also be implemented with any other suitable non-NN-based module.

[0051] In one example embodiment, the NN-based module may be configured to process the reconstructed visual data based on a local feature of the reconstructed visual data. By way of example, the local feature may represent local details of the reconstructed visual data, such as texture variation w ithin a small region, pixel gradients between adjacent pixels and a current pixel, and / or the like. In another example embodiment, the NN-based module may be configured to process the reconstructed visual data based on a global feature of the reconstructed visual data. By way of example, the global feature may represent global information of the reconstructed visual data, such as statistical information (such as mean, variance or the like) of all pixels of the reconstructed visual data In a further example embodiment, the NN-based 22 F1257165PCTmodule may be configured to process the reconstructed visual data based on the local feature and the global feature of the reconstructed 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 Details of processing the reconstructed visual data will be described with reference to Fig. 18 below.

[0052] At 1706, the processed visual data is encoded into a second bitstream of the visual data. By way of example rather than limitation, the encoding may be performed with an encoder corresponding the above-mentioned second codec. In addition, at least one of the reconstructing at 1702 or the encoding at 1706 is performed based on an NN-based coding scheme. The NN-based coding scheme may be implemented with an NN-based coding model As used herein, an NN-based coding model may be a model based on neural network technologies. For example, an NN-based coding 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 coding model described here are merely illustrative and therefore should not be construed as limiting the present disclosure in any' way.

[0053] In view of the above, during the transcoding process involving a bitstream generated according to an NN-based coding scheme, after the visual data is reconstructed from the first bitstream, the reconstructed visual data is processed with the NN-based module, and then the processed visual data is encoded into the second bitstream. Compared with the conventional solution where the reconstructed visual data is directly encoded into a further bitstream, the proposed method can advantageously utilize the processing with NN-based module to optimize the reconstructed visual data during the transcoding process, and thus bridging the gap between the two coding schemes involved in the transcoding process Thereby, the transcoding quality can be improved

[0054] In some embodiments, the reconstructing at 1702 may' be performed based on a non -NN-based coding scheme, and the encoding at 1706 may be performed based on the NN-based coding scheme. By way of example rather than limitation, the non-NN-based coding scheme may comprise Joint Photographic Experts Group (JPEG), JPEG 2000, High Efficiency Image File Format (HEIF), Advanced Video Coding (A VC) intra coding. High Efficiency Video Coding (HEVC) intra coding. Versatile Video Coding (VVC) intra coding, AVC coding, HEVC coding, VVC coding, or the like. The NN-based coding scheme may comprise JPEG Artificial Intelligence (Al) learning-based image coding or any other suitable NN- based coding scheme.

[0055] The inventor has recognized that JPEG relies on the fixed quantization tables and Discrete Cosine Transform (DCT)-based compression, while HE VC uses hand-crafted intra-prediction and quantization, both lacking optimal rate-distortion optimization. These limitations of non-NN-based coding scheme introduce artifacts such as blocking and blurring especially at high compression rates. When rccomprcssing these images, NN-based coding scheme may misinterpret these distortions as meaningful features, leading to significant loss in performance. In aid of processing the reconstructed visual data with the NN-based module, the gap between the non-NN-based coding scheme and NN-based codingF1257165PCTscheme can be bridged and the reconstructed visual data can be adapted for NN-based coding scheme, and thus the coding performance can be improved.

[0056] Now, details about how to process the reconstructed visual data will be discussed with reference to Fig. 18. Fig. 18 illustrates a flowchart of a method 1800 for processing the reconstructed visual data in accordance with some embodiments of the present disclosure. The method 1800 may be implemented at 1704 of Fig. 17.

[0057] At 1802, a first luma feature representation and a first chroma feature representation are generated from the reconstructed visual data. For example, as shown in Fig. 16A, a luma component of the reconstructed visual data may be processed with a convolution, to generate the first luma feature representation (e.g., referring to luminance in Fig. 16A). A chroma component of the reconstructed visual data may be processed with a convolution to generate the first chroma feature representation (e g., referring to chrominance in Fig. 16A). Alternatively, the luma component and chroma component of the reconstructed visual data may be processed together with a stack of convolutions to obtain the first luma feature representation and the first chroma feature representation. It should be understood that the above illustrations are described merely for purpose of description. The scope of the present disclosure is not limited in this respect.

[0058] At 1804, a second luma feature representation is generated by processing the first luma feature representation and the first chroma feature representation with a first branch of the NN-based module.

[0059] Fig. 16A shows a detailed architecture of the ELGC-TransNet in Fig. 15, which may be a non¬ limiting example of the NN-based module in accordance with some embodiments of the present disclosure. As shown in Fig. 16A, there are two branches in the architecture of the ELGC-TransNet. The first branch (e.g, stacked local-global modulation module (LGMM) in Fig. 16A) of the NN-based module (e g, ELGC-TransNet m Fig 16A) may be used for luminance, and the second branch (e g, stacked dynamic fusion module (DFM) in Fig 16A) of the NN-based module (e g, ELGC-TransNet in Fig 16A) may be used for chrominance These two branches composed by stacked LGMM and DFM are used to extract and fuse both global and local information, which may be discussed in detail later.

[0060] In order to generate the second luma feature representation, the first luma feature representation and the first chroma feature representation may be processed with a stack of first modules (e.g., a stack ot’LGMMs in Fig. 16A) of the first branch, to generate an intermediate luma feature representation The generated intermediate luma feature representation and the first luma feature representation is used to generate the second luma feature representation.

[0061] In some embodiments, each first module among the stack of first modules may use a luma feature representation and the first chroma feature representation, as an input, to generate an output luma feature representation.

[0062] In order to generate the output luma feature representation of the first module, a feature channel of the luma feature representation may be extended with a convolution. The extended luma feature representation may be divided into a first luma feature part and a second luma feature part through channel split. The first luma feature part and the first chroma feature representation (e.g., Fc in Fig. 16B) may be processed with a first branch (e.g., MVGG branch in Fig 16B) of the first module to generate 24 F1257165PCTthe global feature. The second luma feature part and the first chroma feature representation (e g., Fc in Fig. 16B) may be processed with a second branch (e g., LFA branch in Fig 16B) of the first module to generate the local feature The global feature and the local feature may be fused, to generate the output luma feature representation of the first module[0063JFor example, the first luma feature part and the first chroma feature representation (e g., Fc in Fig.16B) may be concatenated to obtain a first concatenated feature (e.g., Fx in Fig. 16B). Based on statistical information of the first concatenated feature and a result of processing the first concatenated feature with a stack of a max pooling operation and a deep-wise convolution, an aggregated feature may be generated. Based on the first concatenated feature and a result of processing the aggregated feature with a stack of a convolution, an activation function and an upsampling operation, the global feature may be generated. For example, the statistical information may include mean, variance, standard deviation, or the like. Additionally or alternatively, the activation function may be a Gaussian Error Linear Unit (GELU), a Rectified Linear Unit (RELU), or the like. Additionally or alternatively, the upsampling operation may be a bilinear upsampling operation, an up-shuffle operation or the like.

[0064] In addition, the second luma feature part and the first chroma feature representation may be concatenated to obtain a second concatenated feature(e.g., Fy in Fig. 16B). A reference local feature may be generated by processing the second concatenated feature with at least one convolution, and the local feature may be generated based on the reference local feature In one example embodiment, the at least one convolution may include a plurality of convolutions with different kernel sizes, and the reference local feature may be generated by concatenating respective results of processing the second concatenated feature with each of the plurality of convolutions. Alternatively, the at least one convolution may include only one convolution and the reference local feature may be generated by processing the second concatenated feature with the convolution

[0065] In one example embodiment, the generation of the local feature may be achieved by the LFA branch as shown in Fig 16B. As shown in Fig. 16B. LFA branch may be used for generating of the local feature (e.g., Fi.i in Fig. 16B). For example, the second luma feature part and the first chroma feature representation (e.g, Fc in Fig. 16B) may be concatenated to obtain a second concatenated feature The second concatenated feature may be processed with at least one convolution to generate a reference local feature (e.g., FH in Fig. 16B) By way of example, the at least one convolution may be three convolutions, e.g., 3x3 deep-wise convolution, 5x5 deep-wise convolution, 7x7 deep-wise convolution as shown in Fig.16B. Alternatively, the at least one convolution used to generate the reference local feature (e g, FH in Fig. 16B). may be one convolution

[0066] In some embodiments, the reference local feature may be used to obtain the local feature For example, the local feature may be obtained by processing the reference local feature with a stack of at least one convolution and an activation function. By way of example rather than limitation, the at least one convolution may be two convolution of size 1 x 1. In some embodiments, the activation function may a GELU, an RELU or the like. For example, as shown in Fig. 16B, one convolution of size 1×1, one GELU, and one convolution of size 1 x 1 may be successively stacked. In some alternative embodiments, the reference local feature itself may be directly set as the local feature.F1257165PCT

[0067] Referring back to Fig. 18, at 1806, a second chroma feature representation is generated by processing the first luma feature representation and the first chroma feature representation with a second branch of the NN-based module

[0068] In some embodiments, an intermediate chroma feature representation may be generated by processing the first luma feature representation and the first chroma feature representation with a stack of second modules of the second branch of the NN-based module. This will be described in detail below. The second chroma feature representation may be generated based on the intermediate chroma feature representation and the first chroma feature representation For example, the intermediate chroma feature representation and the first chroma feature representation may be added based on element-wise addition to obtain the second chroma feature representation. Alternatively, the intermediate chroma feature representation and the first chroma feature representation may be fused with a convolution to obtain the second chroma feature representation.

[0069] For example, a stack of the second modules (e.g., the stack of DFMs in Fig. 16A) of the second branch of the NN-based module may be used to process the first luma feature representation and the first chroma feature representation so as to generate an intermediate chroma feature representation. In some embodiments, the first luma feature representation and the first chroma feature representation may be concatenated (e.g., through channel concatenation in Fig. 16A, or the like) to obtain a concatenated feature representation. The concatenated feature representation may be processed with the stack of second modules to generate the intermediate chroma feature representation.

[0070] As shown in Fig. 16A, each second module among the stack of second modules may generate an output chroma feature representation. In some embodiments, the output chroma feature representation may be generated by: extending a feature channel of an input feature representation of the second module with a convolution: dividing the extended input feature representation into a first chroma feature part and a second chroma feature part through channel split; generating a first fusion feature and a second fusion feature based on the first chroma feature part and the second chroma feature part; and generating the output chroma feature representation based on the first fusion feature and the second fusion feature.

[0071] In some embodiments, in order to generate of the first fusion feature and the second fusion feature, for example, the first chroma feature part may be processed with an average pooling operation to generate a smooth feature. The second chroma feature part may be processed with a max pooling operation to generate a salient feature. The salient feature and a result of processing the smooth feature with a stack of a convolution and an activation function may be used to generate the first fusion feature. In addition, the smooth feature and a result of processing the salient feature with a stack of a convolution and an activation function may be used to generate the second fusion feature. For example, the convolution may be a deep-wise convolution or the like. The activation function may a GELU, a RELU or the like.

[0072] In order to generate of the output chroma feature representation, the first fusion feature and the second fusion feature may be concatenated (e.g, with a channel concatenation in Fig 16C or the like) to obtain a third concatenated feature The third concatenated feature may be processed with a stack of an upsampling operation, an activation function and a convolution, to generate the output chroma feature representation. By way of example rather than limitation, the upsampling operation may be a nearestF1257165PCTupsampling operation, an up-shuffle operation or the like., The activation function may include a RELU, a GELU or the like.

[0073] Referring back to Fig. 18, at 1808, the processed visual data is generated based on the second luma feature representation and the second chroma feature representation Tn one example embodiment, the second luma feature representation may be processed with a convolution to generate a processed luma component. The second chroma feature representation may be processed with a convolution to generate a processed chroma component. A result of merging the processed luma component and the processed chroma component may be transformed to a red-green-blue (RGB) domain, to obtain intermediate visual data. The obtained intermediate visual data and the reconstructed visual data may be added (e g., through element-wise addition or the like) to generate the processed visual data. In an alternative embodiment, the second luma feature representation and the second chroma feature representation may be processed with a convolution to obtain the processed visual data.

[0074] While the various steps in the flowchart 1800 are presented and described sequentially, one of ordinary skill in the art will appreciate that some or all of the steps may be executed in different orders and some or all of the steps may be executed in parallel. Further, in various embodiments, one or more of the steps described below may be omitted, repeated, and / or performed in a different order. Accordingly, the specific arrangement of steps shown in Fig 18 should not be construed as limiting the scope of the embodiment.

[0075] For purpose of illustration, the proposed solution will be discussed in more detail according to a non-limiting example with reference to Figs. 16A-16C.

[0076] The overall architecture of the proposed Efficient Local-Global Collaboration (ELGC)-Transcoding is illustrated in Figs. 16A-16C, which incorporates YUV decoupling and local-global collaborative processing Specifically, the decoded RGB image is first converted into luminance and chrominance components in the YUV space A 3 x 3 convolution layer is then applied to extract shallow features for luminance (luma for short) and chrominance (chroma for short), respectively For luminance reconstruction, the extracted shallow features are led into a series of local -global modulation modules (LGMM) to generate deep representative features using the shallow chrominance features as auxiliary information. Each LGMM includes a mean-variance guided global (MVGG) branch and a local feature aggregation (LFA) branch. For chrominance refinement, shallow chrominance features are concatenated with shallow luminance features and fed into a series of dynamic fusion module (DFM), which employs adaptive average pooling and max pooling to extract and fuse complementary features. Next, residual connection is employed to combine shallow and deep features, generating the reconstructed luminance and chrominance components through additional convolution layers, which are finally merged and transformed back to the RGB domain. In order to learn the high-frequency information, a global residual connection may be inserted before the final output.[0077JI11 a local-global modulation module (LGMM), the shallow chrominance feature may be used as auxiliary and propose the LGMM to extract the local details and global information. It mainly contains the mean-variance guided global (MVGG) and the local feature extraction (LFA) branches.

[0078] Given the shallow luminance feature FL, first, the feature channel may be extended using a lx 127 F1257165PCTconvolution and divide it into two parts, which are concatenated with the shallow chrominance feature Fc to get the feature Fx and Fy. Fx and Fy then may be fed into the MVGG and LFA branches in parallel

[0079] Tn the MVGG branch, a max pooling operation is used to capture larger receptive fields and 3x3 deep-wise convolution is performed to obtain global structure representation F ss = DWConv3 (Maxpool ( x)) (4) where DWConv3( • ) represents the 3x3 deep-wise convolution. Maxpool( ■ ) denotes the max pooling operation.

[0080] Furthermore, the global mean and variance of the input feature Fx may be incorporated as statistical information and may be combined with Fs to get the aggregated feature Fx.Fy = Fs +,u(Fx) + F(Fx) (5) where p(Fx) and σ(Fx) denote the mean and variance of Fx. This statistical modulation help exploit the global information of luminance.

[0081] Subsequently, aggregated feature Fx is further enhanced and modulated with the input features Fx to get the global feature FLG, which is denoted by:FLG - FX 0 B inear(GELU(Convi i(Fy))) (6) where Conv1×1( • ) refers to the 1 x 1 convolution, GELU( • ) represents the GELU activation function. Bilinear( • ) denotes the Bilinear upsampling, and © is element-wise multiplication.

[0082] In the LFA branch, considering high compressed images usually lose more detailed information, multiple deep-wise convolutional kernels may be used to extract different receptive field features focusing on the local structures.FH = Concat(DWConv3(Fy), DWConv5(Fy), DWConv7(Fy)) (7) where DWConv3( • ), DWConv5( • ), DWConv7( • ) denote the 3x3, 5x5, 7×7 deep-wise convolution, Concat represents the channel concatenation and Fn is the reference local feature.

[0083] In addition, the FH may be fed into the convolution and GELU activation function to get the local feature FLL. Finally the global feature FLG and local feature FLI. may be merged through a 1 x 1 convolutional layer to obtain This process can be described as:FL = Conv1×1(FLG + FLL) (8) where FLdenotes the output luminance feature.

[0084] In the dynamic fusion module (DFM), due to the fact that decoupled chrominance mainly contains color information, different global features for chrominance refinement may be extracted in a dynamic fusion manner Specifically, the shallow luminance feature Fi. may be taken as the supplemental information and concatenate it with shallow chrominance feature FC to get the feature FCG. Similar to LGMM, FCG is fed into a 1×1 convolution and divide it into two parts used to extract diverse global features.F¹CG, F²CG = SConv1×1(FCG) (9) where S denotes the channel splitting operation.28 F1257165PCT

[0085] Then, F¹CG undergoes an average pooling to obtain smooth feature FA, and F²CG undergoes a max pooling to obtain salient feature FM. Next, FA passes through 3x3 deep-wise convolution, GELU. and pixel-wise multiplication with FM to obtain the enhanced fusion feature F¹CG. Similarly, FM undergoes feature modulation operation to obtain the enhanced fusion feature F2CG.0 GELU(Com’’xi(F.i)) (10) F²CG = FA ⊙ GELU(Conv1×1(FM)) (11)

[0086] Subsequently. F]ccand F‘‘ccare concatenated and fed into the Nearest upsampling, RELU activation function and 1 x 1 convolution to get final refined feature FCG.FCG = Conv1×1(RELU(Nearest(concat(F¹CG, F²CG)))) (12)

[0087] In some embodiments, to align with the objective evaluation metrics of JPEG Al, the YU VDistortion loss is introduced to specifically optimize image quality in the YUV color space. This loss function balances distortions across luminance (Y) and chrominance (UV) components, ensuring compatibility with JPEG Al’s emphasis on YUV-based metrics. Specifically, distortions are computed using Mean Squared Error (MSE) and Multi-Scale Structural Similarity (MS-SSIM) between the raw image and the transcoded image, addressing both computational accuracy and perceptual quality.L = LY + LUV (13) LY = α·LYMSE + β·LYMS-SSIM (14)LUV = γ·LUMSE + γ·LVMSE ( 15)(16) where LY, LU and LV denote the loss for luminance and chrominance components respectively. Specifically, LYMSE and LYMS-SSIM represent the MSE and MS-SSIM metrics of luminance component, while LLUV = γ·LUMSE + γ·LVMSE and LVMSE correspond to the MSE metric of chrominance components. Additionally, a, p and y are hyperparameters that control the weighting of these terms in the loss function.

[0088] According to still further embodiments of the present disclosure, an apparatus for visual data transcoding is provided The apparatus may include a processor and a non -transitory memory with instructions thereon. The instructions upon execution by’ the processor, cause the processor to perform the method as discussed above according to any of the embodiments or examples

[0089] It should be understood that the possible implementations of the NN-based module described with reference to Figs. 16A-16C are merely illustrative and therefore should not be construed as limiting the present, disclosure in any way.

[0090] In view of the above, the solutions in accordance with some embodiments of the present disclosure can advantageously improve the compatibility and performance of visual data transcoding and further extend the application scenarios of NN-based visual data coding scheme.

[0091] According to further embodiments 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 transcoding The method comprises: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least oneF1257165PCTof the reconstructing or the encoding being performed based on an NN-based coding scheme

[0092] According to still further embodiments of the present disclosure, method for storing a bitstream of visual data is provided The method comprises: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN -based coding scheme; and storing the second bitstream in a non -transitory computer-readable recording medium.

[0093] Fig. 19 illustrates a flowchart of a method 1900 for visual data transcoding in accordance with some embodiments of the present disclosure. The method 1900 starts at 1902, where visual data is reconstructed from a first bitstream of the visual data. Furthermore, at 1904, the reconstructed visual data is encoded into a second bitstream of the visual data. At least one of the reconstructing or the encoding is performed with a neural network (NN)-based model, 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.

[0094] In one example embodiment, the reconstructing at 1902 is performed with a first NN-based model (e.g., an NN-based decoder based on JPEG Al image coding standard), and the encoding at 1904 is performed with a non-NN-based encoder, such as an encoder based on JPEG, JPEG 2000, H.264, H.265 or H.266 coding standard. In another example embodiment, the reconstructing at 1902 is performed with a non-NN-based decoder (such as a decoder based on JPEG, JPEG 2000, H 264, H 265 or H 266 coding standard), and the encoding at 1904 is performed with a second NN-based model, e g, an NN-based encoder based on JPEG Al image coding standard. In a further example embodiment, both the reconstructing at 1902 and the encoding at 1904 is performed with an NN-based model. In this case, this NN-based model may, for example, comprise the above-mentioned first NN-based model and the second NN-based model. It should be understood that the above examples are described merely for purpose of description. The scope of the present disclosure is not limited in this respect.

[0095] In view of the above, the reconstructing step and / or the encoding step involved in the visual data transcoding process is performed with an NN-based model. Compared with the conventional solution where the NN-based visual data coding scheme is not supported in transcoding, the proposed method can advantageously support NN-based visual data coding scheme, so as to improve the compatibility of visual data transcoding. Thereby, the application scenarios of NN-based visual data coding scheme is also extended.

[0096] In some embodiments, at 1904, the reconstructed visual data may be processed with an intermediate-processing module. By way of example, the intermediate-processing module may comprise a filtering process, or the like. In addition, the filtering process may be NN-based Then, the processed reconstructed visual data may be encoded into the second bitstream with the second NN-based model. In 30 F1257165PCTthis case, weights of the second NN -based model may be not changed. For example, the weights of the second NN-based model for visual data encoding may be directly used in this transcoding scenario.

[0097] In some embodiments, the second NN-based model may comprise a plurality sets of weights. Each set in the plurality sets of weights may be configured for a specific purpose For example, a first set of weights among the plurality sets of weights may be used for a purpose different from transcoding, such as coding, quality enhancement, bitrate adjustment, or the like. Moreover, a second set of weights among the plurality sets of weights may be used for transcoding. For example, the second set of weights may be dependent on a first codec for the first bitstream. In addition, the bitstream may comprise an indication indicating one of the plurality sets of weights that is used. In some embodiments, the plurality sets of weights may be associated with different architectures, different numbers of layers, different types of NN layers, and / or the like.

[0098] In some embodiments, the visual data may be decoded from the second bitstream For example, the encoding at 1904 may be performed with a second NN-based model, and the decoding may be performed with a third NN-based model.

[0099] In some embodiments, the first bitstream and / or the second bitstream may comprise an indication indicating whether a transcoder is used. In some embodiments, a plurality of encoders may be available for the encoding at 1904, and the first bitstream and / or the second bitstream may comprise an indication indicating one of the plurality of encoders that is used.

[0100] In some embodiments, a post-processing process may be applied on the decoded visual data. Alternatively, information regarding at least one of the following depends on a type of the reconstructed visual data: whether to apply a post -processing process on the decoded visual data, or how to apply the post-processing process on the decoded visual data. In some embodiments, if the first bitstream coded with a specific codec, the post -processing process is applied on the decoded visual data

[0101] In some embodiments, the post-processing process may comprise a filtering process and / or the like. By way of example, the post -processing process may be NN-based.

[0102] In some embodiments, the second bitstream may comprise an indication indicating a type of the reconstructed visual data. For example, the type of the reconstructed visual data may be allowed to be raw visual data. Additionally or alternatively, the type of the reconstructed visual data may be allowed to be visual data decode from the first bitstream coded with a first codec. In this case, the second bitstream further may comprise an indication indicating a kind of the first codec.

[0103] In some embodiments, the second bitstream may comprise a first set of indications indicating parameters of a decoding model for decoding the visual data from the second bitstream. By way of example, the first set of indications may comprise an indication indicating a coefficient of a processing layer of the decoding model.

[0104] In some embodiments, the first bitstream and / or the second bitstream may comprise a second set of indication indicating an entropy coding mode to be used. For example, the second set of indication may comprise: an indication indicating a table used in entropy decoding, an indication indicating an initialization scheme for entropy decoding, and / or the like.

[0105] In some embodiments, the first bitstream and / or the second bitstream may comprise an indication 31 F1257165PCTindicating a synthesis transform to be used.

[0106] In some embodiments, if a first indication in the second bitstream indicates that the second bitstream is obtained without transcoding, a first synthesis transform may be used for processing the second bitstream If the first indication indicates that the second bitstream is obtained by transcoding, a second synthesis transform different from the first synthesis transform may be used for processing the second bitstream.

[0107] In some embodiments, the first bitstream and / or the second bitstream may comprise an indication indicating a color transform. In addition, or alternatively, the first bitstream and / or the second bitstream may comprise an indication indicating an inverse color transform.

[0108] According to further embodiments 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 transcoding. The method comprises: reconstructing visual data from a first bitstream of the visual data; and encoding the reconstructed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed with a neural network (NN) -based model.

[0109] According to still further embodiments of the present disclosure, a method for storing bitstream of visual data is provided. The method comprises: reconstructing visual data from a first bitstream of the visual data; encoding the reconstructed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed with a neural network (NN) -based model; and storing the second bitstream in a non-transitory computer-readable recording medium.

[0110] Although the method 1700 in Fig 17 and the method 1900 in Fig 19 are described separately. One of ordinary skill in the art will appreciate that the processing process with the NN -based module may also be introduced into the method 1900 The scope of the present disclosure is not limited in this respect

[0111] In some embodiments, first information regarding at least one of the following may be indicated in the bitstream: whether to apply the method, or how to apply the method. For example, the first information may be indicated at a block level, a sequence level, a group of pictures level, a picture level, a slice level, a tile group level, and / or the like.

[0112] In some embodiments, the first information may be indicated in one of the following: a coding structure of a coding tree unit (CTU), a coding structure of a coding unit (CU), a coding structure of a transform unit (TU), a coding structure of a prediction unit (PU), a coding structure of a coding tree block (CTB), a coding structure of a coding block (CB), a coding structure of a transform block (TB), a coding structure of a prediction block (PB), a sequence header, a picture header, a sequence parameter set (SPS), a video parameter set (VPS), a dependency parameter set (DPS), a decoding capability information (DCI), a picture parameter set (PPS), an adaptation parameter sets (APS), a slice header, or a tile group header.

[0113] In some embodiments, the first information may be dependent on coded information of the visual data. By way of example rather than limitation, the coded information may comprise a block size, a color format, a single tree partitioning, a dual tree partitioning, a color component, a slice type, a picture type, and / or the like32 F1257165PCT

[0114] In some embodiments, any of the above-mentioned indication may be a syntax element. For example, the syntax element may be binarized as one of the following: a flag, a fixed length code, an exponential Golomb (EG) code, a unaiy code, a truncated unary code, or a truncated binary code In addition, the syntax element may be coded with at least one context model Alternatively, the syntax element may be bypass coded. In some embodiments, the syntax element may be signaled based on a condition.

[0115] In some embodiments, the syntax element may be indicated at one of the following: a block level, a sequence level, a group of pictures level, a picture level, a slice level, or a tile group level. In some embodiments, the syntax element may be indicated in one of the following: a coding structure of a coding tree unit (CTU), a coding structure of a coding unit (CU), a coding structure of a transform unit (TU ), a coding structure of a prediction unit (PU), a coding structure of a coding tree block (CTB), a coding structure of a coding block (CB), a coding structure of a transform block (TB), a coding structure of a prediction block (PB), a sequence header, a picture header, a sequence parameter set (SPS), a video parameter set (VPS), a dependency parameter set (DPS), a decoding capability information (DCI), a picture parameter set (PPS), an adaptation parameter sets (APS), a slice header, or a tile group header.

[0116] In view of the above, the solutions in accordance with some embodiments of the present disclosure can advantageously improve the compatibility of visual data transcoding and extend the application scenarios of NN-based visual data coding scheme.

[0117] Implementations of the present disclosure can be described in view of the following clauses, the features of all of the following clauses can be combined in any reasonable manner. For example, a feature(s) from clause section A. and / or a feature(s) from clause section B can be combined with each other in any suitable manner The division of the three clause sections A, and B are merely for clarity, and therefore should not be construed as limiting the present disclosure m any way

[0118] Cl ause Al A method for visual data transcoding, comprising: reconstructing visual data from a first bitstream of the visual data: processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme

[0119] Clause A2. The method of clause Al, wherein the reconstructing is performed based on a non- NN-based coding scheme, and the encoding is performed based on the NN-based coding scheme.

[0120] Clause A3. The method of clause A2, wherein the NN-based coding scheme comprises Joint Photographic Experts Group (JPEG) Artificial Intelligence (Al) learning-based image coding, and the non-NN-based coding scheme comprises one of the following: JPEG, JPEG 2000, High Efficiency Image File Format (HEIF), Advanced Video Coding (A VC) ultra coding. High Efficiency Video Coding (HE VC) intra coding, or Versatile Video Coding (VVC) intra coding.

[0121] Clause A4. The method of any of clauses A1-A3, wherein the NN-based module is configured to process the reconstructed visual data based on a local feature and a global feature of the reconstructed visual data.

[0122] Clause A5. The method of any of clauses A1-A4, wherein processing the reconstructed visual data 33 F1257165PCTcomprises: generating a first luma feature representation and a first chroma feature representation from the reconstructed visual data; generating a second luma feature representation by processing the first luma feature representation and the first chroma feature representation with a first branch of the NN-based module; generating a second chroma feature representation by processing the first luma feature representation and the first chroma feature representation with a second branch of the NN -based module: and generating the processed visual data based on the second luma feature representation and the second chroma feature representation.

[0123] Clause A.6. The method of clause A5, wherein generating the second luma feature representation comprises: generating an intermediate luma feature representation by processing the first luma feature representation and the first chroma feature representation with a stack of first modules of the first branch; and generating the second luma feature representation based on the intermediate luma feature representation and the first luma feature representation.

[0124] Clause A7. The method of clause A6, wherein an input of each first module among the stack of first modules comprises a luma feature representation and the first chroma feature representation, and an output luma feature representation of the first module is generated by: extending a feature channel of the luma feature representation with a convolution; dividing the extended luma feature representation into a first luma feature part and a second luma feature part through channel split; generating a global feature by processing the first luma feature part and the first chroma feature representation with a first branch of the first module; generating a local feature by processing the second luma feature part and the first chroma feature representation with a second branch of the first module; and generating the output luma feature representation of the first module by fusing the global feature and the local feature.

[0125] Clause A8. The method of clause A7, wherein generating the global feature comprises: concatenating the first luma feature part and the first chroma feature representation to obtain a first concatenated feature; generating an aggregated feature based on statistical information of the first concatenated feature and a result of processing the first concatenated feature with a stack of a max pooling operation and a deep-wise convolution; and generating the global feature based on the first concatenated feature and a result of processing the aggregated feature with a stack of a convolution, an activation function and an upsampling operation.

[0126] Clause A9. The method of clause A8, wherein the statistical information comprises at least one of mean or variance, and / or wherein the activation function is a Gaussian Error Linear Unit (GELU), and / or wherein the upsampling operation is a bilinear upsampling operation.

[0127] Clause A 10 The method of any of clauses A7-A9, wherein generating the local feature comprises: concatenating the second luma feature part and the first chroma feature representation to obtain a second concatenated feature; generating a reference local feature by processing the second concatenated feature with at least one convolution; and obtaining the local feature based on the reference local feature

[0128] Clause Al l The method of clause A10, wherein the at least one convolution comprises a plurality of convolutions with different kernel sizes, and the reference local feature is generated by concatenating respective results of processing the second concatenated feature with each of the plurality of convolutions.34 F1257165PCT

[0129] Clause A12. The method of any of clauses A10-A11, wherein the local feature is obtained by processing the reference local feature with a stack of at least one convolution and an activation function

[0130] Clause A13 The method of clause A12, wherein the activation function is a GELU

[0131] Clause A14 The method of any of clauses A5-A13, wherein generating the second chroma feature representation comprises: generating an intermediate chroma feature representation by processing the first luma feature representation and the first chroma feature representation with a stack of second modules of the second branch of the NN-based module; and generating the second chroma feature representation based on the intermediate chroma feature representation and the first chroma feature representation.

[0132] CTause A15. The method of clause A14, wherein generating the intermediate chroma feature representation comprises: concatenating the first luma feature representation and the first chroma feature representation to obtain a concatenated feature representation; and generating the intermediate chroma feature representation by processing the concatenated feature representation with the stack of second modules.

[0133] Clause A16. The method of clause A15, wherein an output chroma feature representation of each second module among the stack of second modules is generated by: extending a feature channel of an input feature representation of the second module with a convolution; dividing the extended input feature representation into a first chroma feature part and a second chroma feature part through channel split; generating a first fusion feature and a second fusion feature based on the first chroma feature part and the second chroma feature part; and generating the output chroma feature representation based on the first fusion feature and the second fusion feature.

[0134] Clause A17. The method of clause A16, wherein generating the first fusion feature and the second fusion feature comprises: generating a smooth feature by processing the first chroma feature part with an average pool ing operation; generating a sal ient feature by processing the second chroma feature part with a max pooling operation; generating the first fusion feature based on the salient feature and a result of processing the smooth feature with a stack of a convolution and an activation function; and generating the second fusion feature based on the smooth feature and a result of processing the salient feature with a stack of a convolution and an activation function.

[0135] Clause A18. The method of clause A17. wherein the convolution is a deep-wise convolution and / or the activation function is a GELU.

[0136] Clause A19. The method of any of clauses A 16-Al 8, wherein generating the output chroma feature representation comprises: concatenating the first fusion feature and the second fusion feature to obtain a third concatenated feature; and generating the output chroma feature representation by processing the third concatenated feature with a stack of an upsampling operation, an activation function and a convolution

[0137] Clause A20 The method of clause A19, wherein the upsampling operation is a nearest upsampling operation, and / or the activation function comprises a Rectified Linear Unit (RELU).

[0138] Clause A21. The method of any of clauses A5-A20, wherein generating the first luma feature representation and the first chroma feature representation from the reconstructed visual data comprises:35 F1257165PCTgenerating the first luma feature representation by processing a luma component of the reconstructed visual data with a convolution; and generating the first chroma feature representation by processing a chroma component of the reconstructed visual data with a convolution

[0139] Clause A22 The method of any of clauses A5-A21. wherein generating the processed visual data comprises: generating a processed luma component by processing the second luma feature representation with a convolution; generating a processed chroma component by processing the second chroma feature representation with a convolution; obtaining intermediate visual data by transforming a result of merging the processed luma component and the processed chroma component to a red-green-blue (RGB) domain; and generating the processed visual data by adding the intermediate visual data and the reconstructed visual data.

[0140] Clause A23. The method of any of clauses A1-A22, wherein the visual data comprise a video, a picture of the video, or an image

[0141] Clause A24. An apparatus for visual data transcoding 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 A1-A23.

[0142] Clause A25. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of clauses A1-A23.

[0143] Clause A26. 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 transcoding, wherein the method comprises: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN) -based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme

[0144] Cl ause A27 A method for storing a bitstream of visual data, comprising: reconstructing visual data from a first bitstream of the visual data; processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; and encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

[0145] Clause B1. A method for visual data transcoding, comprising: reconstructing visual data from a first bitstream of the visual data: and encoding the reconstructed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed with a neural network (NN)-based model.

[0146] Clause B2. The method of clause Bl, wherein the reconstructing is performed with a first NN-based model.

[0147] Clause B3. The method of any of clauses B 1 -B2, wherein the encoding is performed with a second NN-based model.

[0148] Clause B4. The method of clause B3, wherein encoding the reconstructed visual data comprises: processing the reconstructed visual data with an intermediate -processing module: and encoding the processed reconstructed visual data into the second bitstream with the second NN-based model.36 F1257165PCT

[0149] Clause B5. The method of clause B4, wherein weights of the second NN-based model are not changed

[0150] Clause B6 The method of any of clauses B4-B5, wherein the intermediate-processing module comprises a filtering process

[0151] Clause B7 The method of clause B6, wherein the filtering process is NN-based.

[0152] Clause B8. The method of any of clauses B3-B7, wherein the second NN-based model comprises a plurality sets of weights.

[0153] Clause B9. The method of clause B8, wherein a first set of weights among the plurality sets of weights is used for a purpose different from transcoding.

[0154] Clause B10. The method of any of clauses B8-B9, wherein a second set of weights among the plurality sets of weights is used for transcoding

[0155] Clause B11. The method of clause B10, wherein the second set of weights is dependent on a first codec for the first bitstream.

[0156] Clause B 12 The method of any of clauses B8-B11, wherein the bitstream comprises an indication indicating one of the plurality sets of weights that is used.

[0157] Clause B13. The method of any of clauses B8-B12, wherein the plurality sets of weights are associated with at least one of the following: different architectures, different numbers of layers, or different types of NN layers.

[0158] Clause B14 The method of any of clauses B1-B13, further comprising: decoding the visual data from the second bitstream.

[0159] Clause Bl 5. The method of clause B14, wherein the encoding is performed with a second NN-based model, and the decoding is performed with a third NN-based model

[0160] Clause B16 The method of any of clauses B14-B15, wherein at least one of the first bitstream or the second bitstream comprises an indication indicating whether a transcoder is used

[0161] Clause B17 The method of any of clauses B14-B16, wherein a plurality of encoders are available for the encoding, and at least one of the first bitstream or the second bitstream comprises an indication indicating one of the plurality of encoders that is used.

[0162] Clause B18 The method of any of clauses B 14 -Bl 7, wherein a post-processing process is applied on the decoded visual data

[0163] Clause B19. The method of any of clauses B14-B18, wherein information regarding at least one of the following depends on a type of the reconstructed visual data: whether to apply a post -processing process on the decoded visual data, or how to apply the post -processing process on the decoded visual data.

[0164] Clause B20 The method of clause B19, wherein if the first bitstream coded with a specific codec, the post-processing process is applied on the decoded visual data.

[0165] Clause B21. The method of any of clauses B18-B20, wherein the post-processing process comprises a filtering process.

[0166] Clause B22 The method of any of clauses B18-B21, wherein the post-processing process is NN-based37 F1257165PCT

[0167] Clause B23. The method of any of clauses B1 -B22, wherein the second bitstream comprises an indication indicating a type of the reconstructed visual data.

[0168] Clause B24 The method of clause B23, wherein the type is allowed to be raw visual data

[0169] Clause B25 The method of any of clauses B23-B24, wherein the type is allowed to be visual data decode from the first bitstream coded with a first codec

[0170] Clause B26. The method of clause B25, wherein the second bitstream further comprises an indication indicating a kind of the first codec

[0171] Clause B27 The method of any of clauses B 1 -B26. wherein the second bitstream comprises a first set of indications indicating parameters of a decoding model for decoding the visual data from the second bitstream.

[0172] Clause B28. The method of clause B27, wherein the first set of indications comprises an indication indicating a coefficient of a processing layer of the decoding model.

[0173] Clause B29. The method of any of clauses Bl -B28, wherein at least one of the first bitstream or the second bitstream comprises a second set of indication indicating an entropy coding mode to be used.

[0174] Clause B30. The method of clause B29, wherein the second set of indication comprises at least one of the following: an indication indicating a table used in entropy decoding, or an indication indicating an initialization scheme for entropy decoding.

[0175] Clause B31 The method of any of clauses Bl -B30, wherein at least one of the first bitstream or the second bitstream comprises an indication indicating a synthesis transform to be used

[0176] Clause B32. The method of any of clauses Bl -B31, wherein if a first indication in the second bitstream indicates that the second bitstream is obtained without transcoding, a first synthesis transform is used for processing the second bitstream, and if the first indication indicates that the second bitstream is obtained by transcoding, a second synthesis transform different from the first synthesis transform is used for processing the second bitstream

[0177] Clause B33. The method of any of clauses B1-B32. wherein at least one of the first bitstream or the second bitstream comprises an indication indicating a color transform.

[0178] Clause B34 The method of any of clauses B1-B33, wherein at least one of the first bitstream or the second bitstream comprises an indication indicating an inverse color transform

[0179] Clause B35. The method of any of clauses B1-B34, wherein first information regarding at least one of the following is indicated in the bitstream: whether to apply the method, or how to apply the method.

[0180] Clause B36. The method of clause B.35, wherein the first information is indicated at one of the following: a block level, a sequence level, a group of pictures level, a picture level, a slice level, or a tile group level.

[0181] Clause B37. The method of clause B35, wherein the first information is indicated in one of the follow ing: a coding structure of a coding tree unit (CTU), a coding structure of a coding unit (CU), a coding structure of a transform unit (TU), a coding structure of a prediction unit (PU), a coding structure of a coding tree block (CTB), a coding structure of a coding block (CB), a coding structure of a transform block (TB), a coding structure of a prediction block (PB), a sequence header, a picture header, a sequence 38 F1257165PCTparameter set (SPS), a video parameter set (VPS), a dependency parameter set (DPS), a decoding capability information (DCI), a picture parameter set (PPS), an adaptation parameter sets (APS), a slice header, or a tile group header

[0182] Cl ause B38 The method of any of clauses B35-B37, wherein the first information is dependent on coded information of the visual data.

[0183] Clause B39 The method of clause B38, wherein the coded information comprises at least one of the following: a block size, a color format, a single tree partitioning, a dual tree partitioning, a color component, a slice type, or a picture type

[0184] Clause B40. The method of any of clauses B12-B39, wherein an indication comprises a syntax element

[0185] Clause B41. The method of clause B40, wherein the syntax element is binarized as one of the following: a flag, a fixed length code, an exponential Golomb (EG) code, a unary code, a truncated unary code, or a truncated binary code.

[0186] Clause B42 The method of any of clauses B40-B41, wherein the syntax element is coded with at least one context model, or wherein the syntax element is bypass coded.

[0187] Clause B43 The method of any of clauses B40-B42, wherein the syntax element is signaled based on a condition.

[0188] Clause B44. The method of any of clauses B40-B43, wherein the syntax element is indicated at one of the following: a block level, a sequence level, a group of pictures level, a picture level, a slice level, or a tile group level.

[0189] Clause B45. The method of any of clauses B40-B44, wherein the syntax element is indicated in one of the following: a coding structure of a coding tree unit (CTU). a coding structure of a coding unit (CU), a coding structure of a transform unit (TU), a coding structure of a prediction unit (PU), a coding structure of a coding tree block (CTB), a coding structure of a coding block (CB), a coding structure of a transform block (TB). a coding structure of a prediction block (PB), a sequence header, a picture header, a sequence parameter set (SPS), a video parameter set (VPS), a dependency parameter set (DPS), a decoding capability information (DCI), a picture parameter set (PPS), an adaptation parameter sets (APS), a slice header, or a tile group header.

[0190] Clause B46 The method of any of clauses B1-B45, wherein the visual data comprise a video, a picture of the video, or an image

[0191] Clause B47 An apparatus for visual data transcoding 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 B1-B46.

[0192] Clause B48 A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of clauses B1-B46.

[0193] Clause B49 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 transcoding, wherein the method comprises: reconstructing visual data from a first bitstream of the visual data; and encoding the reconstructed visual data into a second bitstream of the visual data, at least one of the reconstructing or 39 F1257165PCTthe encoding being performed with a neural network (NN) -based model.

[0194] Clause B50. A method for storing a bitstream of visual data, comprising: reconstructing visual data from a first bitstream of the visual data: encoding the reconstructed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed with a neural network (NN)-based model; and storing the second bitstream in a non -transitory’ computer-readable recording medium.Example Device

[0195] Fig. 20 illustrates a block diagram of a computing device 2000 in which various embodiments of the present disclosure can be implemented. The computing device 2000 may be implemented as or included m 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)

[0196] It would be appreciated that the computing device 2000 shown in Fig. 20 is merely for purpose of illustration, without suggesting any limitation to the functions and scopes of the embodiments of the present disclosure in any manner.

[0197] As shown in Fig. 20, the computing device 2000 includes a general-purpose computing device 2000. The computing device 2000 may at least comprise one or more processors or processing units 2010, a memory 2020, a storage unit 2030, one or more communication units 2040, one or more input devices 2050, and one or more output devices 2060

[0198] In some embodiments, the computing device 2000 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 2000 can support any type of interface to a user (such as ’‘wearable” circuitry and the like).

[0199] The processing unit 2010 may be a physical or virtual processor and can implement various processes based on programs stored in the memory 2020. 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 2000. The processing unit 2010 may also be referred to as a central processing unit (CPU), a microprocessor, a controller or a microcontroller.

[0200] The computing device 2000 typically includes various computer storage medium. Such medium can be any medium accessible by the computing device 2000, including, but not limited to, volatile and non-volatile medium, or detachable and non-detachable medium The memory 2020 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),40 F1257165PCTor a flash memory), or any combination thereof. The storage unit 2030 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 2000

[0201] The computing device 2000 may further include additional detach able / non -detachable, volatile / non-volatile memory medium. Although not shown in Fig. 20, 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.

[0202] The communication unit 2040 communicates with a further computing device via the communication medium. In addition, the functions of the components in the computing device 2000 can be implemented by a single computing cluster or multiple computing machines that, can communicate via communication connections. Therefore, the computing device 2000 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

[0203] The input device 2050 may be one ormore of a variety of input devices, such as a mouse, keyboard, tracking ball, voice-input device, and the like. The output device 2060 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 2040, the computing device 2000 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 2000, or any devices (such as a network card, a modem and the like) enabling the computing device 2000 to communicate with one or more other computing devices, if required Such communication can be performed via input / output (I / O) interfaces (not shown)

[0204] In some embodiments, instead of being integrated in a single device, some or all components of the computing device 2000 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 embodiments, cloud computing provides comput ing, 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 embodiments, 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.41 F1257165PCT

[0205] The computing device 2000 may be used to implement video encoding / decoding in embodiments of the present disclosure. The memory 2020 may include one or more visual data transcoding modules 2025 having one or more program instructions These modules are accessible and executable by the processing unit 2010 to perform the functionalities of the various embodiments described herein

[0206] In the example embodiments of performing visual data encoding, the input device 2050 may receive visual data as an input 2070 to be encoded The visual data may be processed, for example, by the visual data transcoding module 2025, to generate an encoded bitstream. The encoded bitstream may be provided via the output device 2060 as an output 2080.

[0207] In the example embodiments of performing visual data decoding, the input device 2050 may' receive an encoded bitstream as the input 2070. The encoded bitstream may be processed, for example, by the visual data transcoding module 2025, to generate decoded visual data. The decoded visual data may be provided via the output device 2060 as the output 2080.

[0208] While this disclosure has been particularly shown and described with references to example embodiments 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 embodiments of the present application is not intended to be limiting.42 F1257165PCT

Claims

We Claim:1 A method for visual data transcoding, comprising:reconstructing visual data from a first bitstream of the visual data:processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; andencoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

2. The method of claim 1, wherein the reconstructing is performed based on a non-NN-based coding scheme, and the encoding is performed based on the NN-based coding scheme.

3. The method of claim 2, wherein the NN-based coding scheme comprises Joint Photographic Experts Group (JPEG) Artificial Intelligence (Al) learning-based image coding, andthe non-NN-based coding scheme comprises one of the following: JPEG, JPEG 2000, High Efficiency Image File Format (HEIF), Advanced Video Coding (A VC) intra coding, High Efficiency Video Coding (HEVC) intra coding, or Versatile Video Coding (VVC) intra coding.

4. The method of any of claims 1-3, wherein the NN-based module is configured to process the reconstructed visual data based on a local feature and a global feature of the reconstructed visual data5 The method of any of claims 1-4, wherein processing the reconstructed visual data comprises: generating a first luma feature representation and a first chroma feature representation from the reconstructed visual data:generating a second luma feature representation by processing the first luma feature representation and the first chroma feature representation with a first branch of the NN-based module;generating a second chroma feature representation by processing the first luma feature representation and the first chrom a feature representation with a second branch of the NN-based module: andgenerating the processed visual data based on the second luma feature representation and the second chroma feature representation.

6. The method of claim 5, wherein generating the second luma feature representation comprises: generating an intermediate luma feature representation by processing the first luma feature representation and the first chroma feature representation with a stack of first modules of the first branch; and generating the second luma feature representation based on the intermediate luma feature representation and the first luma feature representation7 The method of claim 6, wherein an input of each first module among the stack of first modules comprises a luma feature representation and the first chroma feature representation, and an output luma feature 43 F1257165PCTrepresentation of the first module is generated by:extending a feature channel of the luma feature representation with a convolution;dividing the extended luma feature representation into a first luma feature part and a second luma feature part through channel split;generating a global feature by processing the first luma feature part and the first chroma feature representation with a first branch of the first module;generating a local feature by processing the second luma feature part and the first chroma feature representation with a second branch of the first, module; andgenerating the output luma feature representation of the first module by fusing the global feature and the local feature.

8. The method of claim 7, wherein generating the global feature comprises:concatenating the first luma feature part and the first chroma feature representation to obtain a first concatenated feature;generating an aggregated feature based on statistical information of the first concatenated feature and a result of processing the first concatenated feature with a stack of a max pooling operation and a deep-wise convolution; andgenerating the global feature based on the first concatenated feature and a result of processing the aggregated feature with a stack of a convolution, an activation function and an upsampling operation,9. The method of claim 8, wherein the statistical information comprises at least one of mean or variance, and / orwherein the activation function is a Gaussian Error Linear Unit (GELU), and / orwherein the upsampling operation is a bilinear upsampling operation10. The method of any of claims 7-9, wherein generating the local feature comprises: concatenating the second luma feature part and the first chroma feature representation to obtain a second concatenated feature;generating a reference local feature by processing the second concatenated feature with at least one convolution; andobtaining the local feature based on the reference local feature.

11. The method of claim 10, wherein the at least one convolution comprises a plurality of convolutions with different kernel sizes, and the reference local feature is generated by concatenating respective results of processing the second concatenated feature with each of the plurality of convolutions12. The method of any of claims 10-11, wherein the local feature is obtained by processing the reference local feature with a stack of at least one convolution and an activation function.44 F1257165PCT13. The method of claim 12. wherein the activation function is a GELU.14 The method of any of claims 5-13, wherein generating the second chroma feature representation comprises:generating an intermediate chroma feature representation by processing the first luma feature representation and the first chroma feature representation with a stack of second modules of the second branch of the NN-based module; andgenerating the second chroma feature representation based on the intermediate chroma feature representation and the first chroma feature representation15. The method of claim 14, wherein generating the intermediate chroma feature representation comprises:concatenating the first luma feature representation and the first chroma feature representation to obtain a concatenated feature representation; andgenerating the intermediate chroma feature representation by processing the concatenated feature representation with the stack of second modules.

16. The method of claim 15, wherein an output chroma feature representation of each second module among the stack of second modules is generated by:extending a feature channel of an input feature representation of the second module with a convolution: dividing the extended input feature representation into a first chroma feature part and a second chroma feature part through channel split;generating a first fusion feature and a second fusion feature based on the first chroma feature part and the second chroma feature part; andgenerating the output chroma feature representation based on the first fusion feature and the second fusion feature.

17. The method of claim 16, wherein generating the first fusion feature and the second fusion feature comprises:generating a smooth feature by processing the first chroma feature part with an average pooling operation; generating a salient feature by processing the second chroma feature part with a max pooling operation; generating the first fusion feature based on the salient feature and a result of processing the smooth feature with a stack of a convolution and an activation function; andgenerating the second fusion feature based on the smooth feature and a result of processing the salient feature with a stack of a convolution and an activation function.

18. The method of claim 17, wherein the convolution is a deep-wise convolution and / or the activation function is a GELU.45 F1257165PCT19. The method of any of claims 16-18, wherein generating the output chroma feature representation comprises:concatenating the first fusion feature and the second fusion feature to obtain a third concatenated feature; andgenerating the output chroma feature representation by processing the third concatenated feature with a stack of an upsampling operation, an activation function and a convolution20. The method of claim 19, wherein the upsampling operation is a nearest upsampling operation, and / or the activation function comprises a Rectified Linear Unit (RELU).

21. The method of any of claims 5-20, wherein generating the first luma feature representation and the first chroma feature representation from the reconstructed visual data comprises:generating the first luma feature representation by processing a luma component of the reconstructed visual data with a convolution; andgenerating the first chroma feature representation by processing a chroma component of the reconstructed visual data with a convolution.

22. The method of any of claims 5-21, wherein generating the processed visual data comprises: generating a processed luma component by processing the second luma feature representation with a convolution;generating a processed chroma component by processing the second chroma feature representation with a convolution;obtaining intermediate visual data by transforming a result of merging the processed luma component and the processed chroma component to a red-green-blue (RGB) domain; andgenerating the processed visual data by adding the intermediate visual data and the reconstructed visual data.

23. The method of any of claims 1 -22, wherein the visual data comprise a video, a picture of the video, or an image.

24. An apparatus for visual data transcoding 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 1-23.

25. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of claims 1-23.

26. 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 transcoding, wherein the method comprises:46 F1257165PCTreconstructing visual data from a first bitstream of the visual data;processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data; andencoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme.

27. A method for storing a bitstream of visual data, comprising:reconstructing visual data from a first bitstream of the visual data;processing the reconstructed visual data with a neural network (NN)-based module to obtain processed visual data;encoding the processed visual data into a second bitstream of the visual data, at least one of the reconstructing or the encoding being performed based on an NN-based coding scheme; andstoring the second bitstream in a non-transitory computer-readable recording medium.47 F1257165PCT

Images