Image coding method for reconstructing brightness and chrominance component representation
By optimizing YUV420 image coding through a feature decoupling-fusion module and a priori guidance module, the problem of unutilized differences between luminance and chrominance components is solved, achieving efficient image coding and improved reconstruction quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-11
- Publication Date
- 2026-03-27
AI Technical Summary
Existing deep learning-based YUV420 image coding algorithms fail to effectively utilize the differences and correlations between luminance and chrominance components, resulting in decreased coding efficiency. Repeated upsampling and downsampling operations lead to information loss and degradation of coding performance.
A feature decoupling-fusion module and a priori guidance module are used to extract and map features for luminance and chrominance components respectively. Through entropy coding and entropy decoding, combined with cross-component feature parameter extraction and priori guidance, the feature decoupling and fusion process is optimized to avoid information loss and remove redundancy.
It improves image coding efficiency, reduces information loss, and enhances image reconstruction quality and coding performance, especially significantly improving rate-distortion performance in low bit-rate scenarios.
Smart Images
Figure CN121750863A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to an image coding method for reconstructing the representation of luminance and chrominance components. Background Technology
[0002] Image coding algorithms are techniques that convert raw image data into a compressed bitstream using specific representations to reduce data size, facilitating image storage, transmission, and recovery. Image coding algorithms are widely used in modern communications, video streaming, and satellite image processing, especially in environments with limited bandwidth or storage space, where they can significantly improve data transmission and storage efficiency.
[0003] With the development of deep learning technology, deep learning-based image coding algorithms have achieved significant advantages in compression efficiency and image quality. By adaptively learning image data features, they can achieve higher compression efficiency while maintaining image quality. Compared with traditional coding methods, deep learning-based image coding algorithms can adaptively extract image features through end-to-end training and more accurately approximate the latent variable distribution in probabilistic modeling, thereby achieving higher reconstruction quality at the same bit rate. For the widely adopted YUV420 format in industry standards, it effectively reduces data size by maintaining full resolution of luminance (Y) and downsampling chrominance (UV) components, while fully utilizing the human eye's greater sensitivity to brightness. Therefore, it has become an important research direction for deep learning image coding methods.
[0004] Currently, one type of deep learning-based image coding algorithm for YUV420 in the existing technology includes two implementation methods.
[0005] The first implementation method encodes the luminance and chrominance components separately. Specifically, it extracts Y and UV features independently and performs probability estimation and entropy coding operations on each, generating separate bitstreams for the luminance and chrominance components. This method avoids the adverse effects of mismatches in spatial scale and distribution between components on network training and inference, preserving their respective distribution characteristics.
[0006] The second approach uses structural alignment to achieve joint processing between components. Typically, a preprocessing module upsamples the chroma component to the same resolution as the luma component, merges the Y, U, and V channels, and then directly inputs them into the same encoding network. After the decoder outputs the three channels, a post-processing module downsamples the chroma component back to its original resolution. This approach promotes component fusion to some extent, helps remove redundant information, alleviates the problem of scale inconsistencies between components, and facilitates direct feature extraction by the neural network.
[0007] The shortcomings of the aforementioned deep learning-based image coding algorithms for YUV420 include: although these algorithms have made special processing for the structural characteristics of YUV420 itself, making it easier for deep learning networks to better encode the YUV420 color format, they all fail to take into account the differences and correlations between luminance and chrominance components, leading to a decrease in image coding efficiency. Specifically, this includes: First, the potential correlation between luminance and chrominance components is not utilized, and there is redundancy in complementary information between a large number of components, resulting in limited coding efficiency. In low bit rate scenarios, this redundancy will be further aggravated, causing a decrease in rate-distortion performance; Second, repeated upsampling and downsampling operations lead to information loss, especially destroying the detail texture of luminance components and the edge consistency of chrominance components. At the same time, directly aligning components ignores their inherent differences in distribution and frequency domain, causing deviations in feature learning and probability estimation, affecting the overall coding performance. Summary of the Invention
[0008] Embodiments of the present invention provide an image coding method for reconstructing the luminance and chrominance component representation, so as to effectively improve the efficiency of image coding.
[0009] To achieve the above objectives, the present invention adopts the following technical solution.
[0010] An image coding method for reconstructing the representation of luminance and chrominance components includes:
[0011] The Y component image and UV component image of the image to be compressed are obtained. The Y component image and UV component image are mapped to the latent space using the feature decoupling-fusion module to obtain the latent representation of the Y component and the latent representation of the UV component after mapping.
[0012] The mapped latent representations of the Y component and the UV component are quantized respectively to obtain the quantized latent representations of the Y component and the UV component.
[0013] The latent representations of the quantized Y component and UV component are subjected to lossless entropy coding based on their frequency of occurrence to obtain compressed Y component bitstream and compressed UV component bitstream. The compressed Y component bitstream and compressed UV component bitstream are then transmitted to the decoding end.
[0014] At the decoding end, the compressed bitstream of the Y component and the compressed bitstream of the UV component are processed by entropy decoding to obtain the potential representation of the quantized Y component and the potential representation of the UV component.
[0015] The prior guidance module maps the quantized latent representations of the Y component and the UV component to continuous numerical content, thereby obtaining the reconstructed latent representations of the Y component and the reconstructed latent representations of the UV component in the latent space.
[0016] The latent representations of the Y component and UV component reconstructed from the latent space are restored to the image content through an inverse transform operation, thereby obtaining the Y luminance component and UV chrominance component of the reconstructed image to be compressed.
[0017] Preferably, the method further includes:
[0018] The feature decoupling-decoupling module includes an upsampling module and a downsampling module. It acquires the Y component image and UV component image of the image to be compressed. During the entropy coding process, the downsampling module performs intra-component feature extraction and compression on the Y component image and UV component image respectively. Then, redundant information between components is removed through feature interaction between components. The feature interaction between components includes feature decoupling and feature fusion to obtain enhanced Y component image features and UV component image features. The spatial resolution of the enhanced Y component image features and UV component image features is then compressed through the convolutional layer in the downsampling module. The enhanced Y component image features and UV component image features are then nonlinearly mapped and contextually modeled through residual blocks. The enhanced Y component image features and UV component image features are mapped to the latent space respectively to obtain the mapped latent representation of the Y component and the mapped latent representation of the UV component.
[0019] During entropy decoding, the upsampling module expands the spatial resolution of the reconstructed latent representations of the Y component and UV component in the latent space. The upsampling module then combines the residual block to restore the reconstructed latent representations of the Y component and UV component in the latent space to the image content, thereby obtaining the reconstructed Y luminance component and UV chrominance component of the image to be compressed.
[0020] Preferably, the step of mapping the latent representations of the quantized Y component and the latent representations of the UV component to continuous numerical content through the prior guidance module to obtain the reconstructed latent representations of the Y component and the reconstructed latent representations of the UV component in the latent space includes:
[0021] The prior guidance module first decodes the latent representation of the quantized Y component, obtains the structural information of the Y component through the cross-component feature parameter extraction layer, captures the structural information of the Y component through the cross-component feature parameter extraction layer, and maps the latent representation of the quantized Y component into continuous numerical content to obtain the reconstructed latent representation of the Y component in the latent space. The structural information of the Y component is used as a priori condition and passed to the reconstruction path of the UV component to correct the probability distribution parameters and decoding features of the UV component, and maps the latent representation of the UV component into continuous numerical content to obtain the reconstructed latent representation of the UV component in the latent space.
[0022] Preferably, the step of quantizing the mapped latent representation of the Y component and the mapped latent representation of the UV component to obtain the quantized latent representation of the Y component and the quantized latent representation of the UV component respectively includes:
[0023] set up This represents the luminance component of the image to be compressed. This represents the chromaticity components of the image to be compressed. This represents the latent representation of the luminance component of the image to be compressed after feature extraction. This represents the latent representation of the chromaticity components of the image to be compressed after feature extraction. , , Indicates the operation of transformation;
[0024] At the encoding end, respectively and The distribution is modeled to obtain the posterior probabilities of the latent representations of the luminance component and the chrominance component, respectively, as shown in the following formula:
[0025]
[0026]
[0027] in,
[0028] Operations representing probability estimation Represents the posterior probability of latent features of the luminance component. Represents the posterior probability of latent features of chrominance components
[0029] At the encoding end, respectively and Quantification process:
[0030]
[0031]
[0032] Indicates quantization operation
[0033] This represents the latent characterization of the quantized Y component. This represents the potential characterization of the quantized UV components.
[0034] Preferably, the method further includes:
[0035] At the decoding end, the latent representation of the Y component of the latent space is obtained through entropy decoding. and the potential representation of the reconstructed UV components ,right and The distribution is modeled again to obtain the prior probabilities of reconstructing the potential representation of the Y component and the potential representation of the UV component, respectively, as shown in the following formula:
[0036]
[0037]
[0038] in, This represents the prior probability of reconstructing the latent features of the Y component. This represents the prior probability of reconstructing the latent features of the UV components;
[0039] During the training process of the feature decoupling-decoupling module and the prior guidance module, the loss function is used to adjust the probability. , , and Constraints are applied, and the constraint formula is expressed as:
[0040]
[0041]
[0042]
[0043] in, The kl divergence value represents the potential characterization of the luminance component. The kl divergence value represents the potential characterization of the chromaticity components. This represents the operation of calculating the KL divergence of two partial distributions. The parameters are set. This represents the operation of calculating the MSE of the original image and the reconstructed image;
[0044] Based on the value of the loss function, the probability... , , and The value is adjusted, and then the parameters of the feature decoupling-decoupling module and the prior guidance module are adjusted to obtain the optimized feature decoupling-decoupling module and the prior guidance module.
[0045] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention solves the serious performance degradation problem of learning-based image coding networks when encoding YUV420 format images. The present invention, through a dual-branch structure of luminance-chrominance decoupling and fusion, maintains the unique characteristics of each component while avoiding information loss caused by simple upsampling / downsampling, thus solving the component mismatch problem.
[0046] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description
[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0048] Figure 1 A schematic diagram illustrating the implementation principle of an image coding method for reconstructing luminance and chrominance component representation provided in an embodiment of the present invention;
[0049] Figure 2 A flowchart illustrating an image coding method for reconstructing luminance and chrominance component representation provided in an embodiment of the present invention;
[0050] Figure 3 This is a schematic diagram illustrating the processing procedure of the downsampling module in a feature decoupling-fusion module provided in an embodiment of the present invention.
[0051] Figure 4 This is a structural diagram of a priori guidance module provided in an embodiment of the present invention. Detailed Implementation
[0052] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0053] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.
[0054] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.
[0055] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.
[0056] This invention proposes an image coding method for reconstructing the representation of luminance and chrominance components. This method achieves efficient image coding based on deep learning for the YUV420 color format by decoupling luminance and chrominance features at multiple levels and using cross-component prior guidance.
[0057] The implementation principle of the image coding method for reconstructing the luminance and chrominance component representation proposed in this invention is as follows: Figure 1 As shown, this method takes the original image to be encoded as input and mainly includes the following functional modules: residual block, feature decoupling-fusion module and prior guidance module.
[0058] The processing flow of the image coding method for reconstructing the luminance and chrominance component representation proposed in this embodiment of the invention is as follows: Figure 2 As shown, the processing steps include the following:
[0059] Step S10: Image Transformation: Obtain the Y component image and UV component image of the image to be compressed, where the Y component represents the luminance component and the UV components represent the chrominance components. A deep network is used to transform and extract features from the Y component image and UV component image, mapping them to a latent space to obtain the mapped latent representations of the Y component and UV components.
[0060] This processing procedure is... Figure 1 The downsampling, residual block, and feature decoupling-fusion module (downsampling module) in the mid-encoding part are completed together.
[0061] Step S20: Quantization: Quantize the latent representations of the Y component and the UV component after mapping to obtain the quantized latent representations of the Y component and the UV component.
[0062] The purpose of this step is to transform continuous numerical values into discrete numerical values, so that infinite possible values can be transformed into finite possible values, which will facilitate subsequent encoding.
[0063] Step S30, Entropy Coding: The latent representations of the quantized Y and UV components are losslessly represented into compressed bitstream information based on their frequency of occurrence, including compressed Y component bitstream and compressed UV component bitstream, which is the content generated after image compression. Then, the compressed bitstream information is transmitted to the decoding end.
[0064] This processing procedure and the quantization procedure are jointly constituted by Figure 1 The feature decoupling-fusion module (downsampling) at the mid-encoding end is complete.
[0065] Step S40, Entropy Decoding: At the decoding end, the latent representation information of the compressed bitstream of the Y component and the compressed bitstream of the UV component are decoded respectively to obtain the latent representation of the quantized Y component and the latent representation of the UV component.
[0066] This step of the process is handled by Figure 1 The prior guidance module for the decoding end is complete.
[0067] Step S50, Dequantization: Map the discrete quantized latent representations of the Y component and the UV component to continuous numerical content to obtain the reconstructed latent representations of the Y component and the UV component.
[0068] This step of the process is handled by Figure 1 The prior guidance module for the decoding end is complete.
[0069] Step S60, Inverse Transform: The latent representations of the reconstructed Y component and UV component in the latent space are restored to the image content, yielding the reconstructed Y luminance component and UV chrominance components. This step is performed by... Figure 1 The mid-decoding end completes feature decoupling-fusion (upsampling), residual block, and upsampling layer.
[0070] In the image transformation and inverse transformation processes described above, under the YUV420 format, the inconsistent sizes of the Y and UV components make them difficult to process together. This necessitates additional scale alignment operations or complete separation of the two parts without information exchange, leading to image information loss or the inability to remove redundancy between the two parts, thus affecting image quality. This invention designs a feature decoupling-fusion module that can perform upsampling and downsampling. During feature extraction, it maintains the original feature scale ratio of the luminance and chrominance components while ensuring sufficient information exchange. Furthermore, it specifically sets up inter-component feature parameter extraction layers and intra-component feature parameter extraction layers for the two components, removing redundant information and further enhancing the information. This results in more compact and accurate extracted features and the final latent representation, facilitating subsequent encoding operations.
[0071] Transformation process:
[0072]
[0073]
[0074] in,
[0075] Represents the brightness component of the image. Represents the chromaticity components of an image
[0076] This represents the latent representation of the brightness component of the image after feature extraction. This represents the latent representation of the chromaticity components of an image after feature extraction.
[0077] Operations representing transformations
[0078] After obtaining the latent representations, it is necessary to model the distributions of the two representations at the encoding end, and obtain the posterior probabilities of the latent representations of the luminance component and the chrominance component, respectively, as shown in the following formula:
[0079]
[0080]
[0081] in,
[0082] Operations representing probability estimation
[0083] Represents the posterior probability of latent features of the luminance component. Represents the posterior probability of latent features of chrominance components
[0084] Quantification process:
[0085]
[0086]
[0087] Indicates quantization operation
[0088] This represents the potential characterization of the quantized luminance component. The latent characterization of the quantized chromaticity components
[0089] The processes of entropy encoding, dequantization, and entropy decoding are usually omitted, and their formulas are not provided.
[0090] After entropy decoding, the two reconstructed latent representations will be obtained at the decoding end. ,
[0091] After obtaining the reconstructed latent representations, it is necessary to model the distributions of the two representations again at the decoding end, and obtain the prior probabilities of the reconstructed luminance component latent representation and the reconstructed chrominance component latent representation respectively, as shown in the following formula:
[0092]
[0093]
[0094] in,
[0095] This represents the prior probability of reconstructing the latent features of the luminance component. Represents the prior probability of reconstructing the latent features of the chromaticity components.
[0096] In the above process, four different probability distributions were obtained respectively. Encoding end: : Posterior probability of latent features of the brightness component : Posterior probability of latent features of chroma components; Decoding end: : Prior probabilities for reconstructing latent features of the brightness component Prior probabilities for reconstructing latent features of chroma components
[0097] During model training, a loss function is needed to constrain these probabilities (the loss function is not reflected in the network structure) to ensure the quality of the reconstructed image. The constraint method uses KL divergence (the smaller the value, the closer the two distributions are). The constraint formula can be expressed as:
[0098]
[0099]
[0100] in,
[0101] The kl divergence value represents the potential characterization of the luminance component. The kl divergence value represents the latent characterization of the chromaticity components.
[0102] This represents the operation of calculating the KL divergence of two partial distributions.
[0103] Then, by adding the two values to the final loss function, both bitrate and reconstructed image quality can be effectively constrained simultaneously.
[0104]
[0105] in,
[0106] These are parameters that need to be set manually.
[0107] This indicates the operation of calculating the MSE of the original image and the reconstructed image.
[0108] Inverse transform process:
[0109]
[0110]
[0111] Indicates the operation of inverse transformation
[0112] Represents the luminance component of the reconstructed image. This represents the chromaticity components of the reconstructed image.
[0113] Based on the value of the loss function, the probability... , , and The value is adjusted, and then the parameters of the feature decoupling-decoupling module and the prior guidance module are adjusted to obtain the optimized feature decoupling-decoupling module and the prior guidance module.
[0114] Figure 3 This diagram illustrates the processing of a downsampling module within a feature decoupling-fusion module according to an embodiment of the present invention. To address the information redundancy issue caused by processing the Y and UV components separately, the present invention uses a feature decoupling-fusion module to selectively extract latent representations for the luminance and chrominance components. The feature decoupling-fusion module includes an upsampling module and a downsampling module, used for upsampling and downsampling operations in the encoder / decoder. During this process, compact latent feature representations are extracted hierarchically and selectively for the luminance and chrominance components, used in subsequent entropy coding operations.
[0115] Before entropy encoding, the distributions of the Y-component and UV-component features need to be estimated separately. Since the estimation is performed at the encoding end, the network can obtain the original image information at this time, which is called posterior estimation. The upsampling module extracts intra-component features from the Y-component and UV-component images respectively, and then removes redundant information between components through inter-component feature interaction. Specifically, feature interaction includes two important parts: feature decoupling and feature fusion. For each component, in order to maintain its own data characteristics and basic features, branch decoupling is adopted, and then the internal information of the component is used for enhancement. At the same time, to prevent the generation of redundancy between components during decoupling, more compact features are extracted through cross-component feature interaction and feature enhancement. This includes decoupling operations that preserve and enhance the features of the component itself, as well as fusion operations that focus on feature interaction and cross-component enhancement. The combination of decoupling and fusion effectively removes redundant information between components. Then, the spatial resolution is compressed through downsampling convolutional layers to form latent features. Next, nonlinear mapping and context modeling are performed through residual blocks to obtain specific representations for the component characteristics. The design of this module avoids the mismatch problem caused by forcibly aligning the luminance and chrominance components, allowing the detailed texture of the luminance component and the color information of the chrominance component to be learned and extracted independently. At the same time, redundant information between components is gradually removed through feature decoupling and fusion across components, providing a more compact input for subsequent entropy coding.
[0116] Entropy Decoding: Figure 4 This is a structural diagram of a priori guidance module provided in an embodiment of the present invention. Since the feature scale of the luminance component is larger than that of the chrominance component, containing richer image information, the luminance component is decoded first during decoding. Then, the features obtained from the luminance component are used as input to the chrominance component decoding process, guiding and assisting the chrominance component in its decoding operation. This not only allows for more complete utilization of the information-rich luminance component but also helps to make the chrominance component decoding more accurate, ultimately improving the image reconstruction quality.
[0117] Furthermore, to fully utilize the semantic and image structural consistency between the luminance and chrominance components, this invention proposes a priori guidance module for entropy coding. The core idea is to use the structural and edge details contained in the luminance component as priors to guide the probabilistic modeling and feature reconstruction of the chrominance component. In specific implementation, the priori guidance module captures the structural information of the luminance component through a cross-component feature parameter extraction layer and passes it as a conditional input to the reconstruction path of the chrominance component, used to correct the probability distribution parameters and decoding features of the chrominance. In this way, the chrominance component, while maintaining its low-frequency distribution, can achieve structural alignment with the luminance component at edges and textures, thereby reducing artifacts such as color bleeding and edge misalignment, improving the estimation accuracy of the entropy model and the overall quality of the reconstructed image.
[0118] The aforementioned components work together to achieve end-to-end image compression and restoration. While maintaining the characteristics of the luminance and chrominance components, they can fully utilize the correlation between the components and fully explore the relationships between the components in the YUV420 color format, thereby achieving more efficient bitrate allocation and significantly improving the overall performance of image encoding and reconstruction quality.
[0119] In summary, the embodiments of the present invention better adapt to the characteristics of YUV420 by taking into account both decoupling and fusion between components.
[0120] The channel separation and UV component downsampling operations in the YUV420 space lead to differences in scale and data distribution between components. Existing methods often directly separate the luminance and chrominance channels completely or input them into the network uniformly, failing to fully consider the scale differences and distribution characteristics of YUV420. This invention solves the component mismatch problem by using a dual-branch structure of luminance-chrominance decoupling and fusion, which preserves the unique characteristics of each component and avoids information loss caused by simple upsampling / downsampling.
[0121] Effectively eliminate cross-component redundancy:
[0122] The feature decoupling-fusion module proposed in this invention can explicitly decouple luminance and chrominance component information during hierarchical feature extraction, and simultaneously fuse the decoupled features to effectively remove redundant information between luminance and chrominance, forming a more compact representation. This module supports parallel extraction of luminance and chrominance features at different levels and achieves information decoupling and effective fusion through an interactive mechanism. This removes cross-component redundant information, improves the compactness and discriminative power of feature representation, provides more efficient input for subsequent entropy coding, and achieves higher compression efficiency.
[0123] Fully explore the relationships between components to improve image coding performance:
[0124] This invention maintains the basic scale ratio of luminance and chrominance components throughout the entire image encoding and decoding process, making it a design fully compatible with the YUV420 color space. In the feature extraction stage, the luminance and chrominance components are hierarchically decoupled to effectively remove redundant information between components. In the entropy coding stage, based on the consistency between components, the luminance component guides the probability estimation of the chrominance component. By considering both the differences and the consistency between components, this approach has significant research value.
[0125] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.
[0126] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.
[0127] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0128] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An image coding method for reconstructing the representation of luminance and chrominance components, characterized in that, include: The Y component image and UV component image of the image to be compressed are obtained. The Y component image and UV component image are mapped to the latent space using the feature decoupling-fusion module to obtain the latent representation of the Y component and the latent representation of the UV component after mapping. The mapped latent representations of the Y component and the UV component are quantized respectively to obtain the quantized latent representations of the Y component and the UV component. The latent representations of the quantized Y component and UV component are subjected to lossless entropy coding based on their frequency of occurrence to obtain compressed Y component bitstream and compressed UV component bitstream. The compressed Y component bitstream and compressed UV component bitstream are then transmitted to the decoding end. At the decoding end, the compressed bitstream of the Y component and the compressed bitstream of the UV component are respectively processed by entropy decoding to obtain the potential representation of the quantized Y component and the potential representation of the UV component. The prior guidance module maps the quantized latent representations of the Y component and the UV component to continuous numerical content, thereby obtaining the reconstructed latent representations of the Y component and the reconstructed latent representations of the UV component in the latent space. The latent representations of the Y component and UV component reconstructed from the latent space are restored to the image content through an inverse transform operation, thereby obtaining the Y luminance component and UV chrominance component of the reconstructed image to be compressed.
2. The method according to claim 1, characterized in that, The method further includes: The feature decoupling-decoupling module includes an upsampling module and a downsampling module. It acquires the Y component image and UV component image of the image to be compressed. During the entropy coding process, the downsampling module performs intra-component feature extraction and compression on the Y component image and UV component image respectively. Then, redundant information between components is removed through feature interaction between components. The feature interaction between components includes feature decoupling and feature fusion to obtain enhanced Y component image features and UV component image features. The spatial resolution of the enhanced Y component image features and UV component image features is then compressed through the convolutional layer in the downsampling module. The enhanced Y component image features and UV component image features are then subjected to nonlinear mapping and context modeling through residual blocks. The enhanced Y component image features and UV component image features are mapped to the latent space respectively to obtain the latent representation of the Y component and the latent representation of the UV component after mapping. During entropy decoding, the upsampling module expands the spatial resolution of the reconstructed latent representations of the Y component and UV component in the latent space. The upsampling module then combines the residual block to restore the reconstructed latent representations of the Y component and UV component in the latent space to the image content, thereby obtaining the reconstructed Y luminance component and UV chrominance component of the image to be compressed.
3. The method according to claim 1, characterized in that, The method of mapping the quantized latent representations of the Y component and the UV component to continuous numerical content through a priori guidance module to obtain the reconstructed latent representations of the Y component and the UV component in the latent space includes: The prior guidance module first decodes the latent representation of the quantized Y component, obtains the structural information of the Y component through the cross-component feature parameter extraction layer, captures the structural information of the Y component through the cross-component feature parameter extraction layer, and maps the latent representation of the quantized Y component into continuous numerical content to obtain the reconstructed latent representation of the Y component in the latent space. The structural information of the Y component is used as a priori condition and passed to the reconstruction path of the UV component to correct the probability distribution parameters and decoding features of the UV component, and maps the latent representation of the UV component into continuous numerical content to obtain the reconstructed latent representation of the UV component in the latent space.
4. The method according to claim 1, 2, or 3, characterized in that, The process of quantizing the mapped latent representations of the Y component and the UV component to obtain quantized latent representations of the Y component and UV components respectively includes: set up This represents the luminance component of the image to be compressed. This represents the chromaticity components of the image to be compressed. This represents the latent representation of the luminance component of the image to be compressed after feature extraction. This represents the latent representation of the chromaticity components of the image to be compressed after feature extraction. , , Indicates the operation of transformation; At the encoding end, respectively and The distribution is modeled to obtain the posterior probabilities of the latent representations of the luminance component and the chrominance component, respectively, as shown in the following formula: ； ； in, Operations representing probability estimation This represents the posterior probability of the latent features of the luminance component. Represents the posterior probability of latent features of the chrominance component; At the encoding end, respectively and Quantification process: ； ； Indicates quantization operation. This represents the latent characterization of the quantized Y component. This represents the potential characterization of the quantized UV components.
5. The method according to claim 4, characterized in that, The method further includes: At the decoding end, the latent representation of the Y component of the latent space is obtained through entropy decoding. and the potential representation of the reconstructed UV components ,right and The distribution is modeled again to obtain the prior probabilities of reconstructing the potential representation of the Y component and the potential representation of the UV component, respectively, as shown in the following formula: ； ； in, This represents the prior probability of reconstructing the latent features of the Y component. This represents the prior probability of reconstructing the latent features of the UV components; During the training process of the feature decoupling-decoupling module and the prior guidance module, the loss function is used to adjust the probability. , , and Constraints are applied, and the constraint formula is expressed as: ； ； ； in, The kl divergence value represents the potential characterization of the luminance component. The kl divergence value represents the potential characterization of the chromaticity components. This represents the operation of calculating the KL divergence of two partial distributions. The parameters are set. This represents the operation of calculating the MSE of the original image and the reconstructed image; Based on the value of the loss function, the probability... , , and The value is adjusted, and then the parameters of the feature decoupling-decoupling module and the prior guidance module are adjusted to obtain the optimized feature decoupling-decoupling module and the prior guidance module.