Substitute Quality Factor Learning in Latent Space for Neural Image Compression

Abstract

A neural image compression using surrogate quality factor learning in a latent space includes receiving a compressed bitstream and a target quality factor indicating a target compression quality; computing a decoded latent representation of the compressed bitstream; computing a reconstructed image based on the decoded latent representation of the compressed bitstream and the target quality factor; computing shared features based on a network forward calculation using shared decoding parameters (SDPs) of one or more layers of a convolutional neural network; computing estimated ADPs for one or more layers of the convolutional neural network based on the shared features, adaptive decoding parameters (ADPs), and the target quality factor; and computing an output tensor based on the estimated ADPs and the shared features in the one or more layers of the convolutional neural network.

Description

Technical Field

[0001] This application is based on and claims priority to U.S. Provisional Patent Application No. 63 / 289,048, filed on December 13, 2021, U.S. Provisional Patent Application No. 63 / 257,005, filed on October 18, 2021, and U.S. Patent Application No. 17 / 949,434, filed on September 21, 2022, the disclosures of which are hereby incorporated by reference in their entireties.

[0002] Apparatuses and methods consistent with exemplary embodiments of the present disclosure relate to surrogate quality factor learning in a latent space for neural image compression.

Background Art

[0003] ISO / IEC MPEG (JTC 1 / SC 29 / WG 11) is actively exploring potential needs for the standardization of future video coding technologies, including advanced neural image and video compression methodologies. ISO / IEC JPEG has established a JPEG-AI group focused on AI-based end-to-end neural image compression (NIC) using neural networks (NNs).

[0004] Previous approaches have shown promising performance, but flexible bitrate control remains a challenging problem for previous NIC methods. Conventionally, it may be necessary to train multiple model instances that individually target the desired trade-off between rate and distortion (the quality of the compressed image). To reconstruct images from different bitrates, all of these multiple model instances may be stored and deployed on the decoder side. Also, since it is difficult to train and store an infinite number of model instances for all possible target bitrates, it is not possible to provide any smooth bitrate control. In previous approaches, multi-rate NICs have been studied where one model instance is trained to achieve compression at multiple predefined bitrates. However, any smooth bitrate control remains an open and unsolved problem. SUMMARY OF THE INVENTION

[0005] According to some embodiments, a method is provided for neural image compression using surrogate quality factor learning in a latent space. The method is executed by at least one processor and includes receiving a compressed bitstream and a target quality factor indicating a target compression quality, calculating a decoded latent representation of the compressed bitstream, and calculating a reconstructed image based on the decoded latent representation of the compressed bitstream and the target quality factor. The step of calculating the reconstructed image includes computing shared features based on a network forward pass using shared decode parameters (SDP) of one or more layers of a convolutional neural network, computing an estimated adaptive decode parameter (ADP) for one or more layers of the convolutional neural network based on the shared features, the adaptive decode parameter, and the target quality factor, and computing an output tensor based on the estimated ADP and the shared features in one or more layers of the convolutional neural network.

[0006] According to an exemplary embodiment, there may be provided an apparatus including at least one memory configured to store computer program code, and at least one processor configured to access the at least one memory and operate as instructed by the computer program code. The computer program code includes reception code configured to cause the at least one processor to receive a compressed bitstream and a target quality factor indicating a target compression quality, first calculation code configured to cause the at least one processor to calculate a decoded latent representation of the compressed bitstream, and second calculation code configured to cause the at least one processor to calculate a reconstructed image based on the decoded latent representation of the compressed bitstream and the target quality factor. The second calculation code includes first computing code configured to cause the at least one processor to compute shared features based on a network forward computation using shared decode parameters (SDP) of one or more layers of a convolutional neural network, second computing code configured to cause the at least one processor to compute an estimated adaptive decode parameter (ADP) for one or more layers of the convolutional neural network based on the shared features, the adaptive decode parameter (ADP), and the target quality factor, and third computing code configured to cause the at least one processor to compute an output tensor based on the estimated ADP and the shared features in one or more layers.

[0007] According to some embodiments, a non-transitory computer-readable recording medium storing instructions may be provided, the instructions, when executed by at least one processor in a decoder, cause the processor to execute a method for neural image compression using surrogate quality factor learning in a latent space, the method comprising receiving a compressed bitstream and a target quality factor indicating a target compression quality, calculating a decoded latent representation of the compressed bitstream, and calculating a reconstructed image based on the decoded latent representation of the compressed bitstream and the target quality factor, the step of calculating the reconstructed image comprising computing shared features based on a network forward pass using shared decoder parameters (SDP) of one or more layers of a convolutional neural network, computing an estimated adaptive decoder parameter (ADP) for one or more layers of the convolutional neural network based on the shared features, an adaptive decoder parameter, and the target quality factor, and computing an output tensor based on the estimated ADP and the shared features.

Brief Description of the Drawings

[0008] The features, advantages, and significance of exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings. In the accompanying drawings, like reference numerals denote like elements.

[0009]

Figure 1

Figure 2

Figure 3A

Figure 3B

Figure 4A

Figure 4B

Figure 4C

Figure 5A

Figure 5B

Figure 6

[0010] The present disclosure describes a method and apparatus for a meta-neural image compression (meta-NIC) framework by finding a substitutional quality factor (QF) within a decoded latent space. A meta-learning mechanism may be used to adaptively compute a surrogate quality control parameter for each image on an encoder based on the decoded latent features of the input image and a target compression quality. The computed quality-adaptive weight parameter may be improved towards better recovery of the target image when the decoder is reconstructing the image, using the surrogate quality control parameter.

[0011] FIG. 1 is a diagram of an environment 100 in which the methods, apparatuses, and systems described herein may be implemented, according to an embodiment.

[0012] As shown in FIG. 1, the environment 100 may include a user device 110, a platform 120, and a network 130. The devices in the environment 100 may be interconnected via a wired connection, a wireless connection, or a combination of wired and wireless connections.

[0013] The user device 110 includes one or more devices capable of receiving, generating, storing, processing, and / or providing information related to the platform 120. For example, the user device 110 may include a computing device (e.g., a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (e.g., a smartphone, a wireless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smartwatch), or a similar device. In some implementations, the user device 110 may receive information from the platform 120 and / or transmit information to the platform 120.

[0014] The platform 120 includes one or more devices, as described elsewhere in this specification. 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 to be modular such that software components can be swapped in or out. As such, the platform 120 may be easily and / or quickly reconfigured for different uses.

[0015] In some implementations, as shown, platform 120 may be hosted in a cloud computing environment 122. In particular, the implementations described herein will be described as hosting platform 120 in cloud computing environment 122, but in some implementations, platform 120 may not be cloud-based (i.e., may be implemented outside of a cloud computing environment) or may be partially cloud-based.

[0016] Cloud computing environment 122 includes an environment that hosts platform 120. Cloud computing environment 122 may provide services such as computing, software, data access, storage, etc. that do not require knowledge of the physical location and configuration of the system and / or device hosting platform 120 and the end user (e.g., user device 100). As shown, cloud computing environment 122 may include a group of computing resources 124 (collectively referred to as "computing resources 124" or individually as "computing resources 124").

[0017] 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, computing resources 124 may host platform 120. Cloud resources may include computing instances running on computing resources 124, storage devices provided on computing resources 124, data transfer devices provided by computing resources 124, and the like. In some implementations, computing resources 124 may communicate with other computing resources 124 via a wired connection, a wireless connection, or a combination of wired and wireless connections.

[0018] As further shown in FIG. 1, the computing resource 124 includes a group of cloud resources such as one or more applications ("APP") 124-1, one or more virtual machines ("VM") 124-2, virtualized storage ("VS") 124-3, and one or more hypervisors ("HYP") 124-4.

[0019] 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 may eliminate the need to install and execute software applications on the user device 110. For example, the application 124-1 may include software related to 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 may send information to / receive information from one or more other applications 124-1 via the virtual machine 124-2.

[0020] The virtual machine 124-2 includes a software implementation of a machine (e.g., a computer) that executes programs like a physical machine. The virtual machine 124-2 may be either a system virtual machine or a process virtual machine depending on the degree of use and correspondence to any physical machine by the virtual machine 124-2. The system virtual machine may provide a complete system platform that supports the execution of a complete operating system ("OS"). The process virtual machine may execute a single program and support a single process. In some implementations, the virtual machine 124-2 may execute on behalf of a user (e.g., the user device 110) and manage the infrastructure of the cloud computing environment 122 such as data management, synchronization, or long-term data transfer.

[0021] The virtualized storage 124-3 includes one or more storage systems and / or one or more devices that use virtualization technology within the memory system or device of the computing resource 124. In some implementations, within the context of a storage system, the types of virtualization may include block virtualization and file virtualization. Block virtualization may refer to the abstraction (or separation) of logical storage from physical storage so that the storage system can be accessed regardless of the physical storage or heterogeneous structure. This separation may enable the administrator of the storage system to have flexibility in how to manage storage for end users. File virtualization may eliminate the dependency between the data accessed at the file level and the location where the file is physically stored. This may enable optimization of the performance of storage usage, server integration, and / or continuous file migration.

[0022] The hypervisor 124-4 may provide hardware virtualization technology that enables multiple operating systems (e.g., "guest operating systems") to run simultaneously on a host computer such as the computer resource 124. The hypervisor 124-4 may present a virtual operating platform to the guest operating system and manage the execution of the guest operating system. Multiple instances of various operating systems may share the virtualized hardware resources.

[0023] Network 130 includes one or more wired and / or wireless networks. For example, network 130 may include a cellular network (e.g., a fifth-generation (5G) network, a long-term evolution (LTE) network, a third-generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., a public switched telephone network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, an optical fiber-based network, etc., and / or combinations of these or other types of networks.

[0024] The number and arrangement of the devices and networks shown in FIG. 1 are provided as an example. In practice, there may be additional devices and / or networks, fewer devices and / or networks, different devices and / or networks, or devices and / or networks with different arrangements than those shown in FIG. 1. Further, two or more devices shown in FIG. 1 may be implemented within a single device, or a single device shown in FIG. 1 may be implemented as a plurality of distributed devices. Additionally or alternatively, a set of devices (e.g., one or more devices) of environment 100 may perform one or more operations described as being performed by another set of devices of environment 100.

[0025] FIG. 2 is a block diagram of exemplary components of one or more devices of FIG. 1.

[0026] Device 200 may correspond to user device 110 and / or platform 120. As shown in FIG. 2, device 200 may include bus 210, processor 220, memory 230, storage component 240, input component 250, output component 260, and communication interface 270.

[0027] Bus 210 includes components that enable communication among the components of device 200. Processor 220 may be implemented in hardware, firmware, or a combination of hardware and software. Processor 220 may be a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field programmable gate array (FPGA), an 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 execute operations. 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) for storing information and / or instructions for use by processor 220.

[0028] Storage component 240 stores information and / or software related to the operation and use of device 200. For example, storage component 240 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optical disk, and / or a solid state disk), a compact disk (CD), a digital versatile disk (DVD), a floppy disk, a cartridge, a magnetic tape, and / or another type of non-transitory computer-readable medium, along with a corresponding drive.

[0029] The input component 250 includes components that enable the device 200 to receive information via user input (e.g., a touch screen display, keyboard, keypad, mouse, button, switch, and / or microphone). Additionally or alternatively, the input component 250 may include sensors for sensing information (e.g., a Global Positioning System (GPS) component, accelerometer, gyroscope, and / or actuator). The output component 260 includes components that provide output information from the device 200 (e.g., a display, speaker, and / or one or more light emitting diodes (LEDs)).

[0030] The communication interface 270 includes transceiver-like components (e.g., a transceiver and / or separate receiver and transmitter) that enable the device 200 to communicate with other devices via a wired connection, wireless connection, or a combination of wired and wireless connections. The communication interface 270 may enable 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, optical interface, coaxial interface, infrared interface, radio frequency (RF) interface, Universal Serial Bus (USB) interface, Wi-Fi® interface, cellular network interface, etc.

[0031] The device 200 may execute one or more of the processes described herein. The device 200 may execute these processes in response to a processor executing 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 spread across multiple physical storage devices.

[0032] Software instructions may be read into memory 230 and / or storage component 240 from another computer-readable medium or from another device via communication interface 270. When executed, the software instructions stored in memory 230 and / or storage component 240 may cause processor 220 to execute one or more processes described herein. Additionally or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to execute one or more processes described herein. Accordingly, embodiments described herein are not limited to any particular combination of hardware circuitry and software.

[0033] The number and arrangement of components shown in FIG. 2 are provided as an example. In practice, device 200 may include additional components, fewer components, different components, or components in a different arrangement than those shown in FIG. 2. Additionally or alternatively, a set of components of device 200 (e.g., one or more components) may perform one or more operations described as being performed by another set of components of device 200.

[0034] The present disclosure proposes a meta-NIC framework that supports surrogate QF in the decoded latent space. Using a meta-learning mechanism, surrogate quality control parameters for each image on the encoder may be adaptively computed based on the decoded latent features of the input image and the target compression quality. The computed quality-adaptive weight parameters may be improved towards better recovery of the target image when the decoder is reconstructing the image using the surrogate quality control parameters.

[0035] An input image x of size (h, w, c) is given, where h, w, and c are the height, width, and number of channels, respectively. The goal of the test stage of the NIC workflow is described as follows. The input image x may be a normal image frame (t = 1), a 4D video sequence including multiple image frames (t > 1), etc. Each image frame may be a color image (c = 3), a grayscale image (c = 1), an rgb + depth image (c = 4), etc. A compressed representation that can be compact for storage and transmission

Number

Number

Number

Number

Number

Number

Number

Number

[0036] Training with a large hyperparameter λ may result in a compression model with less distortion but more bit consumption, and vice versa. Conventionally, for each predefined hyperparameter λ, a NIC model instance is trained, which does not work well for other values of the hyperparameter λ. Thus, to achieve multiple bitrates of a compressed stream, conventional methods may require training and storing multiple model instances. Further, since it is actually difficult to train a model for all possible values of the hyperparameter λ, conventional methods cannot achieve any smooth quality control such as any smooth bitrate control. Additionally, the model instances need to be trained to optimize the loss measured by each type of metric (e.g., for each distortion metric, i.e., PSNR, SSIM, a weighted combination of both, or other metrics), and conventional methods cannot achieve smooth quality metric control.

[0037] FIGS. 3A and 3B are block diagrams of meta-NIC architectures 300A and 300B for adaptive neural image compression by meta-learning according to embodiments.

[0038] As shown in FIG. 3A, the meta-NIC architecture 300A includes a shared decode NN 305 and an adaptive decode NN 310.

[0039] As shown in FIG. 3B, the meta-NIC architecture 300B includes shared decode layers 325 and 330 and adaptive decode layers 335 and 340.

[0040] In the present disclosure, the model parameters of the lower-layer NIC encoder and the lower-layer NIC decoder are separated into two parts, which respectively represent the shared decoding parameter (SDP) and the adaptive decoding parameter (ADP).

Number

Number

[0041] In FIG. 3A, the SDP and the ADP are separate individual NN modules, and these individual modules are sequentially connected to each other for network forward computing. Here, FIG. 3A shows the order in which these individual NN modules are connected. Other orders may be used as well.

[0042] In FIG. 3B, the parameters may be split within the NN layer.

Number

Number

[0043] The embodiment of FIG. 3A may be regarded as one case of FIG. 3B, where the layers in the shared decoding NN315

Number

[0044] In some embodiments, the NN structure of the encoder has no restrictions. For example, for each image x, the compressed representation [Number] may be generated by an NN-based encoder. The compressed representation [Number] is then quantized and entropy encoded to generate a bitstream [Number] which is then entropy decoded and dequantized to generate the decoded latent representation [Number] In some embodiments, individual encoder model instances may be used for each desired compression quality. In other embodiments, a meta-NIC encoder similar to the meta-NIC decoder may be used with shared and adaptive encoding parameters.

[0045] Figures 4A and 4B are block diagrams of an apparatus 400 for adaptive neural image compression by meta-learning during a test stage according to an embodiment. Additionally, FIG. 4C is a block diagram of an inference workflow of a meta-NIC decoder.

[0046] As shown in FIG. 4A, the apparatus 400 includes a decoder 410 and a meta-NIC decoder 420.

[0047] As shown in FIG. 4B, the meta-NIC architecture 400B includes a decoder 410, a Substitutional Perturbation Generation module 420, and a meta-NIC decoder 430.

[0048] In FIG. 4C, the meta-NIC architecture 400B includes an SDP inference module 422, an ADP prediction module 424, and an ADP inference module 426.

[0049] FIG. 4A shows the overall workflow of the decoder in the test stage of the meta-NIC framework.

Number

Number

Number

Number

[0050] In FIGS. 4A and 4B, a compressed bitstream

Number

Number

Number

Number

Number

Number

Number

Number

[0051] FIG. 4C provides an embodiment of the inference workflow of the meta-NIC decoder for the j-th layer. f(j) and

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

[0052] The workflow described in Figure 4C is a general notation.

Number

Number

[0053] Assuming there are a total of M layers for the meta NIC decoder, the output of the last layer may result in a reconstructed image

Number

[0054] Additionally, in Figure 4B, the decoded latent

Number

Number

Number

Number

Number

Number

Number

[0055] Figures 5A and 5B are block diagrams of meta NIC architectures 500A and 500B for the encoder workflow in the test phase by meta learning according to an embodiment.

[0056] As shown in Figure 5A, the meta NIC architecture 500A includes an NN encoding module 505, an encoding module 510, a decoding module 515, a meta NIC decoding module 520, a distortion loss computing module 525, and a backpropagation module 530.

[0057] As shown in Figure 5B, the meta NIC architecture 500B includes an NN encoding module 535, an encoding module 540, a decoding module 545, a substitute perturbation generation module 550, a meta NIC decoding module 555, a distortion loss computing module 560, and a backpropagation module 565.

[0058] In Figure 5A, when the input image x is given and the original target QF

Number

Number

Number

Number

Number

Number

[0059] Here, the weights

Number

Number

Number

[0060] Next, the decoded latent from the decode module 510

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

Number

[0061] In some embodiments, the updated target QF

Number

[0062] In FIG. 5B, the decoded latent

Number

Number

Number

Number

Number

Number

Number

[0063] The proposed Meta NIC framework enables any smooth QF

Number

Number

[0064] for the Meta NIC decoder

Number

Number

Number

[0065] The MetaNIC encoder and the MetaNIC decoder may be trained in an end-to-end manner via a MAML (Model-Agnostic Meta-Learning) mechanism. When the underlying MetaNIC encoder and decoder are trained, the surrogate perturbation generation module is trained by fixing the MetaNIC encoder and decoder parameters while minimizing the fitting loss to obtain a surrogate latent representation

Number

Number

Number

Number

Number

[0066] FIG. 6 is a flowchart of an embodiment of a process 600 for neural image compression using alternative quality factor learning in a latent space.

[0067] As shown in FIG. 6, in operation 610 of process 600, a compressed bitstream and a target compression quality are received. The process proceeds to operation 620 where a decoded latent representation of the compressed bitstream is calculated as shown in FIGS. 4A and 4B. That is, the decoded latent representation may be used to calculate a reconstructed image.

[0068] The process proceeds to operation 630, where the shared features may be computed based on the SDP. The process proceeds to operation 640, where ADP is computed for one or more layers of the convolutional neural network, as shown in FIG. 4C. Thus, the output tensor may be computed based on the estimated ADP and the shared features.

[0069] The foregoing disclosure provides examples and descriptions, but is not intended to be exhaustive and is not intended to limit implementations to the exact forms disclosed. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations. Additionally, one or more features or components of one embodiment may be incorporated into or combined with another embodiment (or one or more features of another embodiment). Further, it is understood that in the flowchart and operation descriptions provided below, one or more operations may be omitted, one or more operations may be added, one or more operations may be executed simultaneously (at least in part), and the order of one or more operations may be switched.

[0070] It will be apparent that the systems and / or methods described herein may be implemented in different forms of hardware, firmware, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and / or methods is not limiting. Thus, the operations and behaviors of the systems and / or methods are described herein without reference to specific software code. It is understood that software and hardware may be designed to implement the systems and / or methods based on the descriptions herein.

[0071] Even if a particular combination of features is defined in the claims and / or disclosed in the specification, these combinations do not limit the disclosure of possible implementations. In fact, many of these features may be combined in ways that are not specifically defined in the claims and / or not disclosed in the specification. Each of the dependent claims listed below may depend directly on only one claim, but the disclosure of possible implementations includes each dependent claim combined with all other claims in the claim set.

[0072] Any element, act, or instruction used in this specification should not be construed as important or essential unless explicitly described. Also, as used in this specification, the articles "a" and "an" are intended to include one or more items and may be used interchangeably with "one or more". When only one item is intended, the term "one" or similar language is used. Also, as used in this specification, terms such as "has", "have", "having", "include", "including", etc. are intended to be open-ended terms. Further, the phrase "based on" is intended to mean "at least partially based on" unless explicitly stated otherwise. Further, expressions such as "at least one of [A] and [B]" or "at least one of [A] or [B]" should be understood to include only A, only B, or both A and B.

Claims

1. A method for neural image compression, executed by at least one processor, comprising: receiving a compressed bitstream and a target quality factor indicating a target compression quality; calculating a decoded latent representation of the compressed bitstream; receiving, by a meta neural image compression (NIC) decoder module including a plurality of sequentially connected layers, the decoded latent representation of the compressed bitstream as an input, and calculating a reconstructed image based on the decoded latent representation and the target quality factor; computing a reconstruction loss between the original input image from which the compressed bitstream was generated and the reconstructed image; computing a gradient of the computed reconstruction loss and updating the target quality factor to minimize the reconstruction loss by adjusting the target quality factor using backpropagation based on the gradient; updating the reconstructed image based on the updated target quality factor and the decoded latent representation of the compressed bitstream; wherein the model parameters of the meta NIC decoder module are separated into a set of shared decoding parameters (SDP) and a set of adaptive decoding parameters (ADP), the SDP set includes a plurality of SDPs respectively associated with the plurality of layers of the meta NIC decoder module, and the ADP set includes a plurality of ADPs respectively associated with the plurality of layers of the meta NIC decoder module; the step of calculating the reconstructed image includes, for each of the plurality of layers of the meta NIC decoder module: (i) computing features by inference based on an input to the layer and an SDP corresponding to the layer; (ii) predicting an updated ADP based on the input to the layer, the features, an ADP corresponding to the layer, and the target quality factor; (iii) computing an output tensor based on the updated ADP and the features; wherein the output tensor output by the last layer of the plurality of layers of the meta NIC decoder module is the reconstructed image.

2. The method according to claim 1, further comprising computing a surrogate latent representation of the compressed bitstream based on the decoded latent representation of the bitstream and the target quality factor.

3. The method according to claim 2, further comprising computing the reconstructed image and updating the target quality factor based on the surrogate latent representation.

4. The method according to claim 3, wherein the reconstruction loss is minimized in response to a determination that the surrogate latent representation is computed based on surrogate perturbation generation.

5. An apparatus comprising: at least one memory configured to store computer program code; at least one processor configured to access the at least one memory and operate as instructed by the computer program code, the computer program code comprising: reception code configured to cause the at least one processor to receive a compressed bitstream and a target quality factor indicating a target compression quality; first computation code configured to cause the at least one processor to compute a decoded latent representation of the compressed bitstream; second computation code configured to input the decoded latent representation of the compressed bitstream to a neural image compression (NIC) decoding module including a plurality of layers connected in sequence, and to compute a reconstructed image based on the decoded latent representation and the target quality factor; fifth computing code configured to cause the at least one processor to compute a reconstruction loss between the original input image from which the compressed bitstream was generated and the reconstructed image, to compute a gradient of the computed reconstruction loss, and to update the target quality factor using backpropagation based on the gradient so that the reconstruction loss is minimized, and to update the reconstructed image based on the updated target quality factor and the decoded latent representation of the compressed bitstream; comprising The model parameters of the meta-NIC decoding module are separated into a set of shared decoding parameters (SDP) and a set of adaptive decoding parameters (ADP). The SDP set includes a plurality of SDPs respectively associated with the plurality of layers of the meta-NIC decoding module. The ADP set includes a plurality of ADPs respectively associated with the plurality of layers of the meta-NIC decoding module. The second computing code is in each of the plurality of layers of the meta-NIC decoding module. (i) First computing code configured to cause the at least one processor to compute a feature amount by inference based on an input to the layer and the SDP corresponding to the layer. (ii) Second computing code configured to cause the at least one processor to predict an updated ADP based on the input to the layer, the feature amount, the ADP corresponding to the layer, and the target quality factor. (iii) Third computing code configured to cause the at least one processor to compute an output tensor based on the updated ADP and the feature amount. The apparatus, wherein the output tensor output by the last layer of the plurality of layers of the meta-NIC decoding module is the reconstructed image. An apparatus, wherein the output tensor output by the last layer of the plurality of layers of the meta-NIC decoding module is the reconstructed image.

6. The apparatus according to claim 5, wherein the computer program code further includes sixth computing code configured to cause the at least one processor to compute a surrogate latent representation of the compressed bitstream based on the decoded latent representation of the bitstream and the target quality factor.

7. The apparatus according to claim 6, wherein the sixth computing code is further configured to cause the at least one processor to compute the reconstructed image and update the target quality factor based on the surrogate latent representation.

8. A non-transitory computer-readable recording medium storing instructions, which when executed by at least one processor, cause the processor to: Receive a compressed bitstream and a target quality factor indicating a target compression quality. Calculating a decoded latent representation of the compressed bitstream; Inputting the decoded latent representation of the compressed bitstream into a neural image compression (NIC) decoding module including a plurality of layers connected in sequence, and calculating a reconstructed image based on the decoded latent representation and the target quality factor; Computing a reconstruction loss between the original input image from which the compressed bitstream was generated and the reconstructed image; Computing a gradient of the computed reconstruction loss, and updating the target quality factor so that the reconstruction loss is minimized by adjusting the target quality factor using backpropagation based on the gradient; Updating the reconstructed image based on the updated target quality factor and the decoded latent representation of the compressed bitstream; performing; The model parameters of the meta-NIC decoding module are separated into a set of shared decoding parameters (SDP) and a set of adaptive decoding parameters (ADP). The SDP set includes a plurality of SDPs respectively associated with the plurality of layers of the meta-NIC decoding module. The ADP set includes a plurality of ADPs respectively associated with the plurality of layers of the meta-NIC decoding module. The step of calculating the reconstructed image is performed in each of the plurality of layers of the meta-NIC decoding module, (i) computing features by inference based on an input to the layer and an SDP corresponding to the layer; (ii) predicting an updated ADP based on the input to the layer, the features, an ADP corresponding to the layer, and the target quality factor; (iii) computing an output tensor based on the updated ADP and the features, A non-transitory computer-readable recording medium, wherein the output tensor output by the last layer of the plurality of layers of the meta-NIC decoding module is the reconstructed image. Claim 9 The non - transitory computer - readable recording medium according to claim 8, wherein the command causes the at least one processor to perform a step of computing a substitute latent representation of the compressed bitstream based on the decoded latent representation of the bitstream and the target quality factor.

10. The non - transitory computer - readable recording medium according to claim 9, wherein the command causes the at least one processor to perform a step of computing the reconstructed image and a step of updating the target quality factor based on the substitute latent representation.

11. The non - transitory computer - readable recording medium according to claim 10, wherein the reconstruction loss is minimized when the substitute latent representation is computed based on substitute perturbation generation.

Images