Method, apparatus, and storage medium for neural image compression using latent feature domain intra prediction

Abstract

A method for decoding an image using latent feature domain intra prediction, executed by at least one processor and including: receiving a set of latent blocks, and for each block in the set of latent blocks: predicting the block based on a previously recovered set of blocks; receiving a selection signal indicating a currently recovered block, and based on the selection signal, performing one of (1) and (2): (1) generating a compact residual, a set of residual context parameters, and a decoded residual; and generating a first decoded block; (2) generating a second decoded block based on the compact representation block and the set of context parameters. The method further includes: generating a set of recovered blocks, the set of recovered blocks including each of the currently recovered blocks; generating a recovered latent image by merging all the blocks in the set of recovered blocks; and decoding the recovered latent image to obtain a reconstructed image.

Description

[0001] Cross - reference to related applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 133,704, filed on January 4, 2021, and to U.S. Patent Application No. 17 / 462,287, filed on August 31, 2021. The disclosures of both of these U.S. applications are hereby incorporated by reference in their entireties. Background Art

[0003] Standard groups and companies have actively searched for potential needs for standardizing future video coding technologies. These standard groups and companies have established the JPEG - AI group, which focuses on AI - based end - to - end neural image compression using deep neural networks (DNNs). The success of recent methods has brought increasing industrial interest in advanced neural image and video compression methods.

[0004] Given an input image x, the goal of NIC is to use the image x as the input to a DNN encoder to compute a compressed representation (the compressed representation is relatively compact for storage and transmission), and then use as the input to a DNN decoder to reconstruct the image Previous NIC methods employed a variational autoencoder (VAE) structure, where the DNN encoder directly used the entire image x as the input to the DNN encoder, and x passed through a set of network layers that worked like a black box to compute an output representation Accordingly, the DNN decoder used the entire representation as the input to the DNN decoder, passed through another set of network layers that worked like another black box to compute the reconstructed

[0005] Block - based intra - prediction and residual coding mechanisms encode the residual between a predicted block and an original block, rather than directly encoding the entire original image. This mechanism has proven to be very effective for compressing image frames with modern video coding standards such as HEVC and VVC. The entire image is partitioned into blocks of various sizes, and a predicted block is generated by replicating the boundary pixels of previously compressed blocks along various angular directions, and then the residual between the original block and the predicted block is compressed. The residual can be encoded more efficiently compared to the original pixels, thus achieving better coding performance. Summary of the Invention

[0006] According to an embodiment, a method for neural image compression using an intra-prediction mechanism in a latent feature domain, executed by at least one processor and including: receiving a set of latent blocks, and for each block in the set of latent blocks: predicting the block using a first neural network based on a previously recovered block set; receiving a selection signal indicating a current recovered block; based on the received selection signal, performing one of (1) and (2): (1) generating a compact residual, a set of residual context parameters, and a decoded residual; and generating a first decoded block based on the predicted block and the decoded residual; (2) generating a second decoded block based on the compact representation block and the set of context parameters. The method further includes: generating a set of recovered blocks, the set of recovered blocks including each of the current recovered blocks; generating a recovered latent image by merging all the blocks in the set of recovered blocks; and decoding the generated recovered latent image using a second neural network to obtain a reconstructed image.

[0007] According to an embodiment, an apparatus for neural image compression using an intra-prediction mechanism in a latent feature domain, including: at least one memory configured to store program code; and at least one processor configured to read the program code and operate according to the instructions of the program code, the program code including: a receiving code configured to cause the at least one processor to receive a set of latent blocks; a prediction code configured to cause the at least one processor to predict a block using a second neural network based on a previously recovered block set; a selection code configured to cause the at least one processor to receive a selection signal indicating a current recovered block, and for each block in the set of latent blocks, based on the received selection signal, perform one of (1) and (2): (1) a first generation code configured to cause the at least one processor to generate a compact residual; a second generation code configured to cause the at least one processor to generate a set of residual context parameters; a third generation code configured to cause the at least one processor to generate a decoded residual; and a first decoding code configured to cause the at least one processor to generate a first decoded block based on the predicted block and the decoded residual; (2) a second decoding code configured to cause the at least one processor to generate a second decoded block based on the compact representation block and the set of context parameters. The program further includes: a recovered block generation code configured to cause the at least one processor to generate a set of recovered blocks, the set of recovered blocks including each of the current recovered blocks; a merging code configured to cause the at least one processor to merge all the blocks in the set of recovered blocks to generate a recovered latent image; and a third decoding code configured to cause the at least one processor to decode the generated recovered latent image using a second neural network to obtain a reconstructed image.

[0008] According to an embodiment, a non-transitory computer-readable medium stores instructions that, when executed by at least one processor, cause: receiving a set of potential blocks; predicting blocks using a second neural network based on a previously recovered block set; receiving a selection signal indicating a currently recovered block, and for each block in the set of potential blocks, based on the received selection signal, performing one of (1) and (2): (1) generating a compact residual, a set of residual context parameters, and a decoded residual; and generating a first decoded block based on the predicted block and the decoded residual; (2) generating a second decoded block based on the compact representation block and the set of context parameters. The non-transitory computer-readable medium further includes instructions that, when executed by at least one processor, cause: generating a recovered block set that includes each of the currently recovered blocks; merging all blocks in the recovered block set to generate a recovered latent image; and decoding the generated recovered latent image using a second neural network to obtain a reconstructed image. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a diagram of an environment that can implement the methods, apparatuses, and systems described herein according to an embodiment.

[0010] Figure 2 is Figure 1 a block diagram of example components of one or more devices.

[0011] Figure 3 is a block diagram of a test NIC encoder and NIC decoder apparatus for performing neural image compression using intra-frame prediction in a latent feature domain during a test phase according to an embodiment.

[0012] Figure 4 is during a test phase according to an embodiment, Figure 3 a block diagram of the decoder side of a test NIC encoder and NIC decoder apparatus.

[0013] Figure 5 is a block diagram of a training apparatus for performing neural image compression using intra-frame prediction in a latent feature domain during a training phase according to an embodiment.

[0014] Figure 6 is a flowchart of a method for performing neural image compression using intra-frame prediction in a latent feature domain according to an embodiment.

[0015] Figure 7 is a block diagram of an apparatus for performing neural image compression using intra-frame prediction in a latent feature domain according to an embodiment. DETAILED DESCRIPTION

[0016] The embodiment relates to a Neural Image Compression (NIC) framework that compresses an input image using a block-based intra prediction mechanism in a latent feature representation through a deep neural network (DNN). Exemplary embodiments of the present invention will be described below with reference to the accompanying drawings. In the drawings, the same modules are denoted by the same reference numerals, and thus repeated descriptions may be omitted as needed. Figure 1 FIG. Figure 1 is a diagram of an environment 100 that can implement the methods, apparatuses, and systems described herein according to an embodiment.

[0017] As Figure 1 shown, the environment 100 may include a user device 110, a platform 120, and a network 130. The devices of the environment 100 may be interconnected via a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection.

[0018] The user device 110 includes one or more devices capable of receiving, generating, storing, processing, and / or providing information associated with the platform 120. For example, the user device 110 may include a computing device (such as a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (such as a smartphone, a radiotelephone, etc.), a wearable device (such as smart glasses or a smart watch), or a similar device. In some implementations, the user device 110 may receive information from the platform 120 and / or send information to the platform 120.

[0019] The platform 120 includes one or more devices as described elsewhere herein. In some implementations, the platform 120 may include a cloud server or a group of cloud servers. In some implementations, the platform 120 may be designed as a modular platform such that software components can be loaded in or out. Thus, the platform 120 can be easily and / or quickly reconfigured for different uses.

[0020] In some implementations, as shown in the figure, the platform 120 may reside in a cloud computing environment 122. It should be noted that although the implementations described herein describe the platform 120 as residing in the cloud computing environment 122, in some implementations, the platform 120 may not be cloud-based (i.e., may be implemented outside of a cloud computing environment) or may be partially cloud-based.

[0021] The cloud computing environment 122 includes an environment that hosts the platform 120. The cloud computing environment 122 may provide services such as computing, software, data access, storage, etc. without the end user (such as the user device 110) knowing the physical location and configuration of the systems and / or devices that host the platform 120. As shown in the figure, the cloud computing environment 122 may include a set of computing resources 124 (collectively referred to as "computing resources 124", and a single computing resource is referred to as "computing resource 124").

[0022] The computing resources 124 include one or more personal computers, workstation computers, server devices, or other types of computing and / or communication devices. In some implementations, the computing resources 124 may control the platform 120. Cloud resources may include computing instances running in the computing resources 124, storage devices provided in the computing resources 124, data transmission devices provided by the computing resources 124, and the like. In some implementations, the computing resources 124 may communicate with other computing resources 124 via a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection.

[0023] As Figure 1 Further shown, the computing resources 124 include a set of cloud resources, such as one or more applications ("APP") 124-1, one or more virtual machines ("VM") 124-2, virtualized memory ("VS") 124-3, one or more hypervisors ("HYP") 124-4, and the like.

[0024] The application 124-1 includes one or more software applications that can be provided to and / or accessed by the user device 110 and / or the platform 120. The application 124-1 can eliminate the need to install and run software applications on the user device 110. For example, the application 124-1 may include software associated with the platform 120 and / or any other software that can be provided via the cloud computing environment 122. In some implementations, one application 124-1 can send information to and / or receive information from one or more other applications 124-1 via the virtual machine 124-2.

[0025] The virtual machine 124-2 includes a software implementation of a machine (e.g., a computer) that runs programs similar to a physical machine. Depending on the degree of correspondence and use of the virtual machine 124-2 to any real machine, the virtual machine 124-2 can be a system virtual machine or a process virtual machine. The system virtual machine can provide a complete system platform that supports the operation of a complete operating system ("OS"). The process virtual machine can run a single program and can support a single process. In some implementations, the virtual machine 124-2 can run on behalf of a user (e.g., the user device 110) and can manage the infrastructure of the cloud computing environment 122, such as data management, synchronization, or long-term data transmission.

[0026] The virtualized memory 124-3 includes one or more storage systems and / or one or more devices that use virtualization technology within the storage system or device of the computing resource 124. In some implementations, in the context of a storage system, the types of virtualization can include block virtualization and file virtualization. Block virtualization can refer to the abstraction (or separation) of logical storage from physical storage, such that the storage system can be accessed without regard to the physical storage or heterogeneous architecture. The separation can allow the administrator of the storage system to have flexibility in how the administrator manages the storage of end users. File virtualization can eliminate the dependency between the data accessed at the file level and the location where the files are physically stored. This can enable optimization of memory usage, server consolidation, and / or non-disruptive file migration performance.

[0027] The hypervisor 124-4 can provide hardware virtualization technology that allows multiple operating systems (such as "guest operating systems") to run simultaneously on a host computer such as the computing resource 124. The hypervisor 124-4 can present a virtual operating platform to the guest operating systems and can manage the operation of the guest operating systems. Multiple instances of various operating systems can share the virtualized hardware resources.

[0028] The network 130 includes one or more wired networks and / or wireless networks. For example, the network 130 can include cellular networks (such as fifth generation (5G) networks, long-term evolution (LTE) networks, third generation (3G) networks, code division multiple access (CDMA) networks, etc.), public land mobile networks (PLMN), local area networks (LAN), wide area networks (WAN), metropolitan area networks (MAN), telephone networks (such as the Public Switched Telephone Network (PSTN)), private networks, ad hoc networks, intranets, the Internet, fiber-based networks, etc., and / or combinations of these networks or other types of networks.

[0029] Figure 1 The number and arrangement of the devices and networks shown are provided as an example. In practice, there can be additional devices and / or networks, fewer devices and / or networks, different devices and / or networks, or devices and / or networks arranged differently from Figure 1 the devices and / or networks shown. Additionally, Figure 1 two or more of the devices shown can be implemented within a single device, or Figure 1The single device shown can be implemented as multiple distributed devices. Additionally or alternatively, a set of devices (e.g., one or more devices) of environment 100 can perform one or more functions described as being performed by another set of devices of environment 100.

[0030] Figure 2 is Figure 1 a block diagram of example components of one or more devices.

[0031] Device 200 can correspond to user device 110 and / or platform 120. As Figure 2 shown, device 200 can include bus 210, processor 220, memory 230, storage component 240, input component 250, output component 260, and communication interface 270.

[0032] Bus 210 includes components that permit communication between the components of device 200. Processor 220 is implemented in hardware, firmware, or a combination of hardware and software. Processor 220 is a central processing unit (CPU), graphics processing unit (GPU), accelerated processing unit, microprocessor, microcontroller, digital signal processor (DSP), field programmable gate array (FPGA), application specific integrated circuit (ASIC), or another type of processing component. In some implementations, processor 220 includes one or more processors that can be programmed to perform functions. Memory 230 includes random access memory (RAM), read only memory (ROM), and / or another type of dynamic or static storage device (e.g., flash memory, magnetic memory, and / or optical memory) that stores information and / or instructions for use by processor 220.

[0033] Storage component 240 stores information and / or software related to the operation and use of device 200. For example, storage component 240 can include a hard disk (e.g., a magnetic disk, optical disk, magneto-optical disk, and / or solid state disk), compact disk (CD), digital versatile disk (DVD), floppy disk, cartridge disk, magnetic tape, and / or another type of non-transitory computer readable medium, as well as corresponding drives.

[0034] Input component 250 includes components that permit device 200 to receive information, such as via user input (e.g., a touch screen display, keyboard, keypad, mouse, button, switch, and / or microphone). Additionally or alternatively, input component 250 can include sensors for sensing information (e.g., a global positioning system (GPS) component, accelerometer, gyroscope, and / or actuator). Output component 260 includes components that provide output information from device 200 (e.g., a display, speaker, and / or one or more light emitting diodes (LEDs)).

[0035] The communication interface 270 includes transceiver-like components (such as a transceiver and / or separate receiver and transmitter), which enable the device 200 to communicate with other devices via, for example, a wired connection, a wireless connection, or a combination of a wired connection and a wireless connection. The communication interface 270 may allow the device 200 to receive information from another device and / or provide information to another device. For example, the communication interface 270 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, etc.

[0036] The device 200 may execute one or more of the processes described herein. The device 200 may execute these processes in response to the processor 220 running software instructions stored by a non-transitory computer-readable medium such as the memory 230 and / or the storage component 240. The computer-readable medium is defined herein as a non-transitory memory device. The memory device includes a memory space within a single physical storage device or a memory space distributed across multiple physical storage devices.

[0037] The software instructions may be read into the memory 230 and / or the storage component 240 from another computer-readable medium or from another device via the communication interface 270. When run, the software instructions stored in the memory 230 and / or the storage component 240 may cause the processor 220 to execute one or more of the processes described herein. Additionally or alternatively, hardwired circuitry may be used in place of or in combination with the software instructions to execute one or more of the processes described herein. Thus, the implementations described herein are not limited to any particular combination of hardware circuitry and software.

[0038] Figure 2 The number and arrangement of the components shown are provided as an example. In practice, the device 200 may include additional components, fewer components, different components, or components arranged differently from those Figure 2 shown. Additionally or alternatively, a set of components (e.g., one or more components) of the device 200 may perform one or more functions described as being performed by another set of components of the device 200.

[0039] Now, methods and apparatuses for NIC via intra-frame prediction and residual coding based on latent feature domain blocks will be described in detail.

[0040] Embodiments may relate to an intra-frame prediction and residual coding framework for NIC based on latent feature domain blocks. Two mechanisms are used to improve the NIC coding efficiency: encoding the residual between the predicted block and the original block instead of encoding the original block, and performing intra-frame prediction in the latent feature domain.

[0041] Figure 3It is a block diagram of a test NIC encoder and NIC decoder device 300 for performing neural image compression using intra prediction in a latent feature domain during a test phase according to an embodiment.

[0042] As Figure 3 shown, the test device 300 includes a DNN main encoding module 301, a partitioning module 302, an intra prediction module 303, a residual Q module 304, a residual AE module 305, a residual context calculation module 306, a Q module 307, an AE module 308, a context calculation module 309, a residual AD module 310, a residual IQ module 311, an AD module 312, an IQ module 313, a block selection module 314, a merging module 315, and a DNN main decoding module 316.

[0043] Given an input image x of size (h, w, c), where h, w, and c are the height, width, and number of channels respectively, the DNN main encoding module 301 calculates a latent representation y by using a DNN main encoder. The latent representation y is a 3D tensor of size (h, w, c), and the latent representation y passes through the partitioning module 302 and is partitioned into n blocks Each partitioned latent block b i has a size (k h , k w , k c ). Suppose represents the set of previously recovered blocks, which passes through the intra prediction module 303 to calculate a predicted block by using a prediction DNN. The prediction residual can be calculated based on the difference between the predicted block i and the partitioned latent block b The prediction residual passes through the residual Q module 304 and is quantized using a quantization method. Then it passes through the residual AE module 305 to generate an entropy - encoded compact residual representation r′ i . Meanwhile, the residual context calculation module 306 calculates a set of residual context parameters z based on the prediction residual r by using a residual context DNN.

[0044] On the other hand, the partitioned latent block b i of the latent representation y can pass through the Q module 307 and then through the AE module 308 to generate (using the quantization method by the Q module 307) a quantized and then (using the arithmetic coding method by the AE module 308) entropy - encoded compact representation b′ i . Meanwhile, the context calculation module 309 is based on the partitioned latent block b i, by using the context DNN to calculate the context parameter set z b .

[0045] Use compact residual representation r′ i and residual context parameters z r , the residual AD module 310 (using arithmetic decoding method), followed by the residual IQ module 311 (using dequantization method) calculates the decoded residual Decoding residual Can be added to the prediction block later To get the decoded block Use the compact representation b′ i and context parameter z b , the AD module 312 (using an arithmetic decoding method) and then the IQ module 313 (using a dequantization method) calculate the decoded block The block selection module 314 generates a selection signal s i , select signal s i Indicates which decoding block or Used as the current recovery block This is done, for example, by setting the selection signal s i The method for generating the selection signal s will be described later. i When using the decoding block When selecting signal s i and the compact residual representation r′ i and residual context parameters z r When using the decoding block When selecting signal s i With the compact representation b′ i and context parameter z b Then, use the current recovery block To restore the previously restored block set Update to the current recovery block set The encoder continues to process the next block b i+1 .

[0046] After all n blocks are recovered, the merging module 315 generates the recovered latent representation by combining all recovered blocks Then, the DNN main decoding module 316 recovers the potential representation based on Compute the reconstructed image by using the DNN main decoder

[0047] Figure 4 According to an embodiment during the testing phase, in particular Figure 3Block diagram of the decoder side of the described NIC encoder and NIC decoder device 300.

[0048] As Figure 4 shown, the decoder side includes an intra prediction module 303, a residual AD module 310, a residual IQ module 311, an AD module 312, an IQ module 313, a merge module 315, and a DNN main decoding module 316.

[0049] On the decoder side, as Figure 4 described, after receiving the block selection signal s i the system selects one of the following methods to calculate the restored block If the selection signal s i is based on the compact residual representation r′ i and the residual context parameter z r indicating that the restored block is from the decoded block, the residual AD module 310 and then the residual IQ module 311 are used to calculate the decoded residual Meanwhile, based on the set of previously restored blocks the intra prediction module 303 calculates the predicted block by using the prediction DNN The decoded residual is added backward to the predicted block to obtain the restored block If the selection signal s i is based on the compact representation b′ i and the context parameter z b indicating that the restored block is from the decoded block, the AD module 312 and then the IQ module 313 are used to calculate the restored block Then, the set of previously restored blocks is updated to the current set of restored blocks using the restored block The decoder continues to decode the next restored block After all blocks are restored, the merge module 315 generates a restored latent representation by combining all the restored blocks

[0050] Then, the DNN main decoding module 316 calculates the reconstructed image based on the restored latent representation using the DNN main decoder

[0051] In a preferred embodiment, the DNN main encoder and the DNN main decoder adopt a VAE structure. The present disclosure does not impose any restrictions on the specific network structures of the DNN main encoder and the DNN main decoder.

[0052] ​The latent representation y can be partitioned in different ways. For example, y is a 3D tensor of size (h, w, c), where h, w, and c are the height, width, and channels of the latent representation. y can be partitioned along the channel axis into (h, w, k c ) blocks (i.e., k h = h, k w = w), partitioned in the height and width dimensions into (k h , k w , c) blocks (i.e., k c = c), partitioned into (k h , k w ) blocks within the height and width dimensions of each channel, or partitioned into regular (k h , k w , k c ) blocks.

[0053] The partitioned blocks can be processed in various orders according to a predetermined scan order, or processed in an order determined adaptively by some scan method. For example, from top to bottom along the height axis, from left to right along the width axis, or from shallow to deep along the channel axis. Once the scan order is determined, the prediction DNN uses the set of previously restored blocks to calculate the current predicted block according to this order The prediction DNN can have different network architectures, and the architecture is usually related to the specific shape based on which the blocks are partitioned. For example, for 2D blocks of size (k h , k w ), the blocks can be processed one after another for each channel, and the set of restored blocks can include blocks from the previous channel and the current channel. Therefore, the prediction DNN can include modules used within the channel spatial relationship and cross-channel relationship. Thus, the present disclosure does not impose any restrictions on the specific network structure of the prediction DNN.

[0054] The context DNN calculates the context parameter z b used by the AD module 312 and the IQ module 313 to calculate the restored block based on the encoded compact representation b′ i In a preferred embodiment, the context parameter z is a set of parameters for calculating the probability density of the partitioned latent blocks b b by a density estimation method. Similarly, the residual context DNN calculates the residual context parameter z i used by the residual AD module 310 and the residual IQ module 311 to calculate the restored residual based on the encoded compact residual representation r′ r In a preferred embodiment, the residual context parameter z i to calculate the restored residual In a preferred embodiment, the residual context parameter z ris to calculate the latent residuals by a density estimation method a set of parameters for the probability density. The present disclosure does not impose any restrictions on a specific density estimation method, the distribution format of latent blocks or latent residuals, or the network structures of the context DNN and the residual context DNN.

[0055] The block selection module generates a selection signal s i by calculating the loss caused by encoding the currently partitioned latent block b i using the compact residual representation r′ i or the compact representation b′ i , and selects the one with less loss. In a preferred embodiment, the R-D loss is used to consider distortion and bit rate:

[0056]

[0057]

[0058] where and measure the distortion (e.g., MSE or SSIM) between the partitioned latent block b i and the reconstructed block i based on the encoded compact residual representation r′ i [[ID=3 and the compact representation b′ respectively. R(r′ i ) and R(b′ i ) calculate the bit rates of the compact residual representation r′ i and the compact representation b′ i respectively. ρ r and ρ b are trade-off hyperparameters.

[0059] The training processes of various DNNs in the embodiments will be described. The goal of the training process is to learn the DNN main encoder, the DNN main decoder, the prediction DNN, the context DNN, and the residual context DNN. Figure 5 is a block diagram of a training apparatus 500 for neural image compression using intra-frame prediction and residual coding based on latent feature domain blocks during a training phase according to an embodiment.

[0060] As Figure 5As shown, the training device 500 includes a DNN main encoding module 301, a partitioning module 302, an intra-frame prediction module 303, a training residual Q module 501, a training residual AE module 502, a residual context calculation module 306, a training Q module 503, a training AE module 504, a context calculation module 309, a training residual AD module 505, a training residual IQ module 506, a training AD module 507, a training IQ module 508, a block selection module 314, a block distortion calculation module 509, a rate calculation module 510, a merging module 315, a DNN main decoding module 316, an input distortion calculation module 511, and a weight update module 512.

[0061] For training, for example, first, the weight coefficients of the above DNN to be learned are initialized by using a pre-trained corresponding DNN model or by setting the DNN model to random numbers. Then, given an input training image x, similar to the test phase, the DNN main encoding module 301 calculates a latent representation y by using the current DNN main encoder. The latent representation y passes through the partitioning module 302 and is partitioned into n blocks Each partitioned latent block b i has a size (k h , k w , k c ). Using the previous recovered block set The intra-frame prediction module 303 calculates a predicted block by using the current prediction DNN Then, based on the predicted block and the partitioned latent block b i the prediction residual is calculated based on the difference between them This prediction residual passes through the training residual Q module 510 and then through the training residual AE module 502 to generate a compact residual representation r' i . At the same time, the residual context calculation module 306 calculates a set of residual context parameters z based on the prediction residual by using the current residual context DNN r .

[0062] On the other hand, the partitioned latent block b of the latent representation y i passes through the training Q module 503 and then through the training AE module 504 to generate a compact representation b' i . At the same time, the context calculation module 309 calculates a set of context parameters z based on the partitioned latent block b i , by using the current context DNN b .

[0063] Using the compact residual representation r' i and the residual context parameter z r, train the residual AD module 505, and then train the residual IQ module 506 to calculate the decoding residual Decoding residual Add it backward to the prediction block To obtain the decoding block In addition, use the compact representation b′ i And the context parameter z b , train the AD module 507, and then train the IQ module 508 to calculate the decoding block The block selection module 314 generates a selection signal s i , the selection signal s i Indicates which decoding block Or Is used as the current recovery block This is achieved, for example, by setting the selection signal s i To binary 0 or 1. Calculate the block distortion loss in the block distortion calculation module 509 To measure the distortion of the recovered latent block Compared with the original partitioned latent block b i , such as traditional MSE, MS - SSIM, or a weighted combination of traditional MSE and MS - SSIM. In addition, the rate loss R(s i , r′ i / b′ i ) can be calculated by the rate calculation module 510 to measure the bit consumption of the compressed representation. When the selection signal s i Uses the decoding block , use the compact residual representation r′ i And the residual context parameter z r To calculate the rate loss R(s i , r′ i ). When using the decoding block , use the residual representation b′ i And the context parameter z b To calculate the rate loss R(s i , b′ i ).

[0064] Then, use the current recovery block To update the previous recovery block set Into the current recovery block set The encoder continues to process the next block b i+1 . After all blocks are recovered, the merging module 315 generates a recovered latent representation by combining all the recovered blocks Then, the DNN main decoding module 316 calculates the reconstructed image based on the recovered latent representation By using the current DNN main decoder Then, in the input distortion calculation module 511 (in Figure 5 The CID module 511 calculates the input distortion loss The final reconstruction quality can be measured using, for example, traditional PSNR, MS-SSIM, or a weighted combination of traditional PSNR and MS-SSIM.

[0065] Given the trade-off hyperparameter λ and the regularization hyperparameter β, the joint RD loss can be calculated as:

[0066]

[0067] Training with a larger trade-off hyperparameter λ results in a compressed model with less distortion but with greater bit consumption, and vice versa. Training with a larger regularization hyperparameter β imposes a large penalty on the block-by-block distortion as an additional constraint. The gradient of the joint RD loss can then be calculated, and the gradient of the joint RD loss is back-propagated through the update weight module 512 to update the weight parameters of the DNN main encoder, DNN main decoder, prediction DNN, context DNN, and residual context DNN. Different DNNs can be updated at different times and at different update rates. In addition, any DNN can be learned separately. For example, some data sets (the same or different from the data sets used in the above training process) can be used to train the prediction DNN, main DNN encoder, and main DNN decoder separately. The above training process can be fixed so that only the weight parameters of the remaining DNNs are updated. Some of the weight parameters in the above training process can also be fine-tuned.

[0068] In a preferred embodiment, the training residual Q module 501, the training residual AE module 502, the training residual AD module 505, and the training residual IQ module 506 are different from their corresponding modules during the testing phase. For example, for training purposes, the training residual Q module 501, the training residual AE module 502, the training residual AD module 505, and the training residual IQ module 506 can be a statistical data sampler to simulate the actual encoding and decoding effects of the residual Q module 304, the residual AE module 305, the residual AD module 310, and the residual IQ module 311. Similarly, the training Q module 503, the training AE module 504, the training AD module 507, and the training IQ module 508 are different from their corresponding modules during the testing phase. For example, for training, the training Q module 503 , the training AE module 504 , the training AD module 507 and the training IQ module 508 may be a statistical data sampler to simulate the actual encoding and decoding effects of the Q module 307 , the AE module 308 , the AD module 312 and the IQ module 313 .

[0069] Figure 6It is a flowchart of a method 600 for neural image compression using intra prediction in a latent feature domain according to an embodiment.

[0070] In some implementations, one or more of the process blocks may be executed by the platform 120 Figure 6 In some implementations, one or more of the process blocks may be executed by another device or group of devices separate from the platform 120 or another device or group of devices including the platform 120 (e.g., the user device 110) Figure 6 Although Figure 6 illustrates exemplary blocks of the method, in some implementations, the method may include additional blocks, fewer blocks, different blocks, or blocks arranged differently from Figure 6 the blocks depicted. Additionally or alternatively, two or more blocks of the method may be executed in parallel.

[0071] As Figure 6 shown, in operation 601, the method 600 includes: generating a latent representation of the input image using a DNN main encoder.

[0072] In operation 602, Figure 6 the method includes: dividing the latent representation into a set of latent blocks.

[0073] In operation 603, Figure 6 the method includes: receiving a selection signal that indicates using a first decoding block or a second decoding block as the current recovery block.

[0074] After operation 603, the method proceeds to operations 604 - 609 and operations 610 - 612. In Figure 6 , operation blocks 604 - 609 and operation blocks 610 - 612 appear to be executed in parallel. However, operation blocks 604 - 609 may be executed before or after operation blocks 610 - 612. The present disclosure is not limited to the ordering of the operation blocks mentioned above.

[0075] In operation 604, Figure 6 the method includes: predicting a block using a prediction DNN based on a set of previously recovered blocks.

[0076] In operation 605, Figure 6 the method includes: calculating a prediction residual.

[0077] In operation 606, Figure 6 the method includes: generating a compact residual, which is a quantized and entropy - coded compact residual representation of the prediction residual.

[0078] In operation 607, Figure 6 the method includes: generating a set of residual context parameters using a residual context DNN based on the generated prediction residual.

[0079] In operation 608, Figure 6 the method includes: decoding and using a dequantization method to generate a decoded residual based on the generated compact residual and the residual context parameter.

[0080] In operation 609, Figure 6 the method includes: generating a first decoded block based on the predicted block 604 and the decoded residual from operation 608.

[0081] In operation 610, Figure 6 the method includes: generating a compact representation by quantizing and entropy coding the blocks in the set of latent blocks partitioned in operation 602.

[0082] In operation 611, Figure 6 the method includes: using a context DNN to generate a set of context parameters based on the blocks in the set of latent blocks partitioned in operation 602.

[0083] In operation 612, Figure 6 the method includes: generating a second decoded block based on the compact representation from operation 610 and the set of context parameters from operation 610.

[0084] In operation 613, Figure 6 the method determines whether the current block is the last block in the set of latent blocks partitioned in operation 602. If the current block is the last block, the last block is processed and the method proceeds to operation 614. If the current block is not the last block, the method repeats operations 604 - 612 for the next block in the set of latent blocks partitioned in operation 602.

[0085] In operation 614, Figure 6 the method includes: generating a set of restored blocks that includes each restored block output from operations 604 - 612.

[0086] In operation 615, Figure 6 the method includes: merging the blocks in the set of restored blocks to generate a restored latent representation of the input image.

[0087] In operation 616, Figure 6 the method includes: decoding the generated restored latent representation using a DNN main decoder to obtain a reconstructed image.

[0088] Figure 7 is a block diagram of an apparatus for neural image compression using intra - prediction in a latent feature domain according to an embodiment.

[0089] As Figure 7As shown, the apparatus includes a latent image generation code 700, a partitioning code 701, a prediction code 702, a selection code 703, a calculation code 704, a first generation code 705, a second generation code 706, a third generation code 707, a first decoding code 708, a compact block generation code 709, a fourth generation code 710, a second decoding code 711, a restored block generation code 712, a merging code 713, and a third decoding code 714.

[0090] The latent image generation code 700 is configured to cause at least one processor to generate a latent representation of an input image using a DNN main encoder.

[0091] The partitioning code 701 is configured to cause at least one processor to partition the latent representation into a set of latent blocks.

[0092] The prediction code 702 is configured to cause at least one processor to predict a block based on a previous set of restored blocks using a prediction DNN.

[0093] The selection code 703 is configured to cause at least one processor to receive a selection signal that indicates using a first decoded block or a second decoded block as a current restored block.

[0094] The calculation code 704 is configured to cause at least one processor to calculate a prediction residual.

[0095] The first generation code 705 is configured to cause at least one processor to generate a compact residual, which is a quantized and entropy-coded compact residual representation of the prediction residual.

[0096] The second generation code 706 is configured to cause at least one processor to generate a set of residual context parameters using a residual context DNN based on the calculated prediction residual.

[0097] The third generation code 707 is configured to cause at least one processor to decode and use a dequantization method to generate a decoded residual based on the generated compact residual and the residual context parameters.

[0098] The first decoding code 708 is configured to cause at least one processor to generate a first decoded block based on the predicted block and the decoded residual.

[0099] The compact block generation code 709 is configured to cause at least one processor to quantize and entropy-encode the blocks in the set of latent blocks to generate a compact representation.

[0100] The fourth generation code 710 is configured to cause at least one processor to generate a set of context parameters using a context DNN based on the blocks in the set of latent blocks.

[0101] The second decoding code 711 is configured such that at least one processor generates a second decoded block based on the compact representation and the context parameter set.

[0102] The recovery block generation code 712 is configured such that at least one processor generates a set of recovery blocks, the set of recovery blocks including each recovery block.

[0103] The merging code 713 is configured such that at least one processor merges the blocks in the set of recovery blocks to generate a recovered latent representation of the input image.

[0104] The third decoding code 714 is configured such that at least one processor uses the DNN main decoder to generate a recovered latent representation to obtain the reconstructed image.

[0105] Although Figure 7 illustrates exemplary blocks of the apparatus, in some implementations, the apparatus may include additional blocks, fewer blocks, different blocks, or blocks arranged differently from Figure 7 the depicted blocks. Additionally or alternatively, two or more blocks of the apparatus may be combined.

[0106] The embodiments describe the idea of using two mechanisms to improve the NIC coding efficiency: encoding the residuals between the predicted block and the original block instead of encoding the original block; and performing prediction in the latent feature domain to conveniently incorporate spatial and cross-channel information for effective prediction and reconstruction. This approach to NIC coding advantageously results in a flexible and general framework that adapts to different intra-frame prediction methods, different neural coding methods, and various types of quality metrics.

[0107] The proposed NIC coding methods can be used alone or can be combined in any order. Furthermore, each method (or embodiment), encoder, and decoder can be implemented by processing circuitry (e.g., one or more processors or one or more integrated circuits). In one example, one or more processors execute a program stored in a non-transitory computer-readable medium.

[0108] The foregoing disclosure provides illustration and description, but is not intended to be exhaustive or to limit the implementations to the precise forms disclosed. Modifications and variations may be made in light of the above disclosure, or may be obtained from practice of the implementations.

[0109] As used herein, the term "component" is intended to be broadly construed as hardware, firmware, or a combination of hardware and software.

[0110] Obviously, the systems and / or methods described herein can be implemented in different forms of hardware, firmware, or a combination of hardware and software. The actual special control hardware or software code used to implement these systems and / or methods does not limit the implementation. Therefore, the operations and behaviors of the systems and / or methods described herein are described without reference to specific software code - it should be understood that the software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0111] Even if combinations of features are recited in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of possible implementations. In fact, many of these features can be combined in ways not specifically recited in the claims and / or disclosed in the specification. Although each dependent claim listed below may only directly depend on one claim, the disclosure of possible implementations includes the combination of each dependent claim with every other claim in the claim set.

[0112] Any element, act, or instruction used herein should not be construed as critical or essential unless so expressly described. Additionally, as used herein, the articles "a" and "an" are intended to include one or more items and can be used interchangeably with "one or more." Additionally, as used herein, the term "set" is intended to include one or more items (e.g., related items, unrelated items, combinations of related and unrelated items, etc.) and can be used interchangeably with "one or more." Where only one item is intended, the term "a" or similar language is used. Additionally, as used herein, the terms "having," "containing," "comprising," or similar terms are intended to be open-ended terms. Additionally, the phrase "based on" is intended to mean "at least partially based on" unless otherwise expressly stated.

Claims

1. A method for neural image compression using intra - prediction in a latent feature domain, the method being executed by at least one processor, and the method comprising: Receiving a set of latent blocks, and for each block in the set of latent blocks: Predicting the block using a first neural network based on a set of previously recovered blocks; Receiving a selection signal indicating the currently recovered block; Based on the received selection signal, performing one of (1) and (2): (1) Generating a compact residual by quantizing and entropy - coding a prediction residual; Generating a set of residual context parameters using a fourth neural network based on the prediction residual; And Generating a decoded residual by decoding and de - quantizing the compact residual and the residual context parameters in the set of residual context parameters; And Generating a first decoded block based on the predicted block and the decoded residual; (2) Generating a second decoded block based on a compact representation block and a set of context parameters; Generating a set of recovered blocks, the set of recovered blocks including each of the currently recovered blocks; Generating a recovered latent image by merging all the blocks in the set of recovered blocks; And Decoding the generated recovered latent image using a second neural network to obtain a reconstructed image.

2. The method according to claim 1, wherein The method further comprises: Generating a latent representation of the input image using a third neural network; and Partitioning the latent representation into the set of latent blocks.

3. The method according to claim 1, wherein The method further comprises: Wherein, the compact residual and the set of residual context parameters are generated simultaneously.

4. The method according to claim 1, wherein, The method further comprises: Generating the compact representation block by quantizing and entropy - coding a current latent block from the set of latent blocks; and Generating the set of context parameters using a fifth neural network based on the current latent block, Wherein, the set of context parameters and the compact representation block are generated simultaneously.

5. The method according to claim 4, wherein The prediction residual is the difference between the predicted block and the current latent block from the set of latent blocks.

6. The method according to any one of claims 1 to 5, wherein The set of recovered blocks is used to update the set of previously recovered blocks.

7. The method according to any one of claims 1 to 5, wherein The method further comprises: The selection signal selects at least one of the first decoded block and the second decoded block as the currently recovered block; When the first decoded block is selected, sending the compact residual, the set of residual context parameters, and the selection signal to a decoder as inputs to the decoder; and When the second decoded block is selected, sending the compact representation block, the set of context parameters, and the selection signal to the decoder as inputs to the decoder.

8. An apparatus for neural image compression using intra - prediction in a latent feature domain, the apparatus comprising: At least one memory configured to store program code; And At least one processor configured to read the program code and operate according to the instructions of the program code, the program code including: A receive code configured to cause the at least one processor to receive a set of latent blocks; A predict code configured to cause the at least one processor to predict a block using a first neural network based on a set of previously recovered blocks; Select code configured to cause the at least one processor to receive a selection signal indicating a current recovery block and, for each block in the set of potential blocks, perform one of (1) and (2) based on the received selection signal: (1) First generation code configured to cause the at least one processor to generate a compact residual by quantizing and entropy encoding a prediction residual; Second generation code configured to cause the at least one processor to generate a set of residual context parameters based on the prediction residual using a fourth neural network; Third generation code configured to cause the at least one processor to generate a decoded residual by decoding and dequantizing the compact residual and the residual context parameters in the set of residual context parameters; and First decoding code configured to cause the at least one processor to generate a first decoded block based on a prediction block and the decoded residual; (2) Second decoding code configured to cause the at least one processor to generate a second decoded block based on a compact representation block and a set of context parameters; Recovery block generation code configured to cause the at least one processor to generate a set of recovery blocks including each of the current recovery blocks; Merge code configured to cause the at least one processor to merge all the blocks in the set of recovery blocks to generate a recovered latent image; and Third decoding code configured to cause the at least one processor to decode the generated recovered latent image using a second neural network to obtain a reconstructed image.

9. The apparatus according to claim 8, wherein, The apparatus further includes: Latent image generation code configured to cause the at least one processor to generate a latent representation of an input image using a third neural network; and Partitioning code configured to cause the at least one processor to partition the latent representation of the input image into the set of potential blocks.

10. The device according to claim 8, wherein, The program code further includes: wherein the compact residual and the set of residual context parameters are generated simultaneously.

11. The device according to claim 8, wherein, The program code further includes: Compact block generation code configured to cause the at least one processor to generate the compact representation block by quantizing and entropy encoding a current potential block from the set of potential blocks; and Fourth generation code configured to cause the at least one processor to generate the set of context parameters based on the current potential block using a fifth neural network, wherein the set of context parameters and the compact representation block are generated simultaneously.

12. The device according to claim 11, wherein, The program code further includes calculation code configured to cause the at least one processor to calculate a difference between the prediction block and the current potential block from the set of potential blocks as the prediction residual.

13. The device according to any one of claims 8 to 12, wherein The program code further includes update code configured to cause the at least one processor to update the previous set of recovery blocks using the set of recovery blocks.

14. The device according to any one of claims 8 to 12, wherein, The selection code is further configured to cause the at least one processor to: select at least one of the first decoded block and the second decoded block as the current recovery block; when the first decoded block is selected, send the compact residual, the set of residual context parameters, and the selection signal to the third decoding code; and When the second decoding block is selected, the compact representation block, the context parameter set, and the selection signal are sent to the third decoding code.

15. A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor for neural image compression using latent feature domain intra prediction, cause the at least one processor to perform the method of any one of claims 1 to 7.

Images