Image block compression and decompression method based on deep learning

Abstract

The invention discloses an image block compression and decompression method based on deep learning, and the method comprises the following steps: S1, generating an importance evaluation matrix for an input image through a space attention network, and dynamically dividing image blocks; s2, dividing the image blocks into three types of blocks according to the importance values; s3, respectively carrying out differential compression on the three types of blocks; s4, packaging the compressed data into a three-dimensional binary structure comprising a header file layer, a data layer and a redundancy layer; and S5, reconstructing the image at a decoding end through a parallel decoding unit, and eliminating splicing artifacts by adopting an edge fusion network. Compared with the prior art, through dynamic compression strategy distribution of semantic perception, a multi-mode compression engine and an optimized bit stream structure, the compression efficiency and the quality of a reconstructed image are remarkably improved, and meanwhile, the coding and decoding performance is optimized.

Description

Technical Field

[0001] The present invention relates to the field of image compression technology, and in particular to an image block compression and decompression method based on deep learning. Background Art

[0002] Current image compression technologies are mainly divided into traditional methods (such as JPEG and JPEG2000) and emerging deep learning-based methods. However, these existing technologies still have many shortcomings:

[0003] First, the main drawbacks of traditional block compression technology (such as JPEG2000 block processing) include:

[0004] 1. Lack of semantic awareness: Fixed compression strategies (such as preset quantization tables) are used, which cannot be dynamically adjusted based on image content. This results in insufficient protection of details in key image areas (such as lesions in medical images), which may result in blurring or information loss after compression.

[0005] 2. Blocking and redundancy: Misalignment between block boundaries and semantic content can easily lead to discontinuous blocking artifacts at block boundaries. Furthermore, relatively high bitrates may still be used for compression in non-critical areas, resulting in storage and bandwidth redundancy.

[0006] Secondly, global deep learning compression methods (such as end-to-end models based on CNN / Transformer, such as the Ballé 2018 architecture) have improved overall compression efficiency, but:

[0007] 1. Single compression ratio limitation: A unified compression strategy is used for the entire image, and the compression ratio cannot be dynamically adjusted based on the importance of the region. This results in significant quality degradation in key areas at high compression ratios, or waste of storage in non-critical areas at low compression ratios.

[0008] 2. Hardware performance bottleneck: When processing high-resolution images (such as 4K images), all image features need to be loaded at once, resulting in extremely high video memory usage (e.g., over 12GB), making it difficult to deploy efficiently on ordinary GPUs.

[0009] 3. Low serial processing efficiency: The decoding process of some models (especially those that rely on autoregressive context) cannot be parallelized, resulting in a significant increase in decoding latency and making it unsuitable for real-time applications.

[0010] Secondly, hybrid compression frameworks (such as ROI detection + traditional encoding such as JPEG / HEVC) attempt to combine the perception capabilities of deep learning with the maturity of traditional encoding, but they have the following drawbacks:

[0011] 1. Inefficient separate architecture: The ROI detection network and traditional encoder usually run independently, and feature extraction and compression coding cannot be jointly optimized, resulting in increased overall encoding and decoding delay.

[0012] 2. Poor feature compatibility: Traditional encoders (such as HEVC) cannot effectively utilize the rich semantic features extracted by deep learning (for example, super-prior information is difficult to be directly compatible with traditional quantization tables), resulting in limited improvement in compression rate.

[0013] 3. Insufficient context modeling: Existing hybrid solutions often have fixed anchor point selection (or region division) strategies for simple and complex regions, which cannot fully adapt to content changes and may result in redundant bitstreams or poor processing of complex textures.

[0014] Summary of common defects of existing technologies:

[0015] 1. Compression strategies are disconnected from semantic understanding: Traditional methods rely on manual features, while global deep learning models lack fine-grained control capabilities, and both fail to achieve effective dynamic coupling between compression strategies and image semantic information.

[0016] 2. Fragmentation of the encoding and decoding process: Especially in hybrid frameworks, feature extraction and compression encoding processes are separated, lacking end-to-end integrated design and optimization, resulting in efficiency loss and performance bottlenecks.

[0017] 3. Unoptimized binary structure: The binary code stream storage structure of existing methods is mostly not specifically designed for the characteristics of block compression. It lacks a flexible layered verification mechanism and dynamic parameter embedding capabilities, resulting in the need to improve encoding and decoding efficiency and robustness.

[0018] Therefore, the industry urgently needs a better way to solve the above problems. Summary of the Invention

[0019] In view of this, the present invention proposes a deep learning-based image block compression and decompression method, which effectively overcomes many shortcomings of the existing technology and achieves significant improvements in compression performance, visual quality, and application flexibility. The technical solution of the present invention is as follows:

[0020] On one hand, the present invention discloses a method for image block compression and decompression based on deep learning, comprising the following steps:

[0021] S1. Generate an importance evaluation matrix for the input image through the spatial attention network and dynamically divide the image into blocks;

[0022] S2, classifying image blocks into three categories according to their importance values;

[0023] S3, perform differential compression on the three types of blocks;

[0024] S4, encapsulating the compressed data into a three-dimensional binary structure including a header layer, a data layer, and a redundancy layer;

[0025] S5. At the decoding end, the image is reconstructed through parallel decoding units, and an edge fusion network is used to eliminate splicing artifacts.

[0026] Specifically, in the step S1:

[0027] The S1 step is specifically as follows:

[0028] S1.1. Process the input image using a feature extraction network based on the backbone network and integrated with the attention module, and output a feature map.

[0029] S1.2. Based on the feature map, generate a block suggestion heat map through a block decision head;

[0030] S1.3. Based on the feature map, the semantic importance value of each potential block position is calculated through multi-task learning to form an importance evaluation matrix, and the image blocks are dynamically divided according to the matrix.

[0031] Specifically, the calculation formula of the importance evaluation matrix is:

[0032]

[0033] Where σ is the Sigmoid function, GAP represents global average pooling, Fi is the feature map of the i-th layer, and m is the scaling factor, which is initially 1.0 and is optimized by backpropagation.

[0034] Specifically, in the step S2, the blocks include key blocks, transition blocks and background blocks.

[0035] Specifically, the key blocks are compressed using the residual dense block U-Net architecture, the transition blocks are compressed using the deformable convolutional compression network, and the background blocks are compressed using the quantization-aware MobileNetV3.

[0036] Specifically, the key block adopts a U-Net architecture containing a 5-level residual dense block and is based on CABAC entropy coding.

[0037] Specifically, the transition block adopts a compression network containing 3 layers of deformable convolution and is dynamically adjusted through quantization step size.

[0038] Specifically, the background block adopts the MobileNetv3 architecture embedded with an asymmetric quantization module and combines RLE and Huffman coding.

[0039] Another aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0040] The advantages of the present invention are as follows:

[0041] 1. Significantly improve compression performance and visual quality: Through semantically aware dynamic block segmentation and a multi-scale, differentiated compression strategy, this method can allocate the bit budget more intelligently and finely. As a result, at the same target bitrate, the reconstructed image can achieve higher objective evaluation metrics (such as PSNR and SSIM). At the same time, subjective visual quality is also improved, image details (especially those in important areas) are more fully preserved, and compression artifacts (such as blocking effects and blurring) are effectively suppressed.

[0042] 2. Enhanced semantic understanding and content adaptability: The deep learning model introduced in this paper can automatically perceive the semantic importance of different image regions, allowing the compression strategy to be dynamically and tightly coupled with the image content. Compared with the fixed compression strategies of traditional methods or the single compression strategies of early global deep learning models, this has greater internal content adaptability and flexibility.

[0043] 3. Optimized Codec Architecture and Efficiency: The end-to-end deep learning model avoids the separation and optimization difficulties inherent in hybrid frameworks between feature extraction and compression encoding modules. The block processing mechanism, combined with a parallel decoding architecture on the decoder side (e.g., utilizing CUDA for parallel processing on the GPU), significantly improves codec processing speed. A specially designed optimized bitstream structure also facilitates fast parsing and error checking, improving overall efficiency.

[0044] 4. Flexibility and Scalability: The block-based processing approach reduces the extremely high hardware requirements (especially video memory) for processing high-resolution images all at once, making this method easier to deploy and apply on hardware platforms with varying performance. The model parameters, block-based strategies, and importance thresholds can all be adjusted to specific application scenarios and requirements, providing excellent flexibility and scalability.

[0045] 5. Compatibility Considerations and Robustness Enhancement: By retaining some JPEG standard marker segments in the bitstream header, this method ensures a certain degree of basic metadata compatibility with legacy decoders. Furthermore, the inclusion of a checksum mechanism and optional redundant layers (such as Reed-Solomon codes) in the bitstream structure enhances the robustness of the compressed data during storage and transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 Schematic diagram of the overall process of the image block compression method based on deep learning in an embodiment of the present invention;

[0048] Figure 2 This is a comparison chart of the effects of the present invention and existing mainstream compression methods. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specific embodiments are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the terms "including" and "having" in the description and claims of the present invention and the above-mentioned figure descriptions and any variations thereof are intended to cover non-exclusive inclusions.

[0051] In the description of the specific embodiments of the present invention, the technical terms "first" and "second" are used only to distinguish different objects and should not be understood to indicate or imply relative importance or implicitly specify the quantity, specific order, or primary and secondary relationship of the technical features indicated. In the description of the embodiments of the present invention, the meaning of "plurality" is more than two, unless otherwise specifically defined.

[0052] References to "embodiments" in this disclosure mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the disclosure. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0053] In the description of the embodiments of the present invention, the term "and / or" is simply a description of the association relationship between associated objects, indicating that three possible relationships exist. For example, A and / or B can represent the following three situations: A exists alone, A and B exists simultaneously, and B exists alone. In addition, the character " / " in this invention generally indicates that the associated objects are in an "or" relationship.

[0054] It should be noted that, for the convenience of description, in the following embodiments, all identical technical features are marked with the same symbols.

[0055] Current image compression technologies are mainly divided into traditional methods (such as JPEG and JPEG2000) and emerging deep learning-based methods. However, these existing technologies still have many shortcomings:

[0056] First, the main drawbacks of traditional block compression technology (such as JPEG2000 block processing) include:

[0057] 1. Lack of semantic awareness: Fixed compression strategies (such as preset quantization tables) are used, which cannot be dynamically adjusted based on image content. This results in insufficient protection of details in key image areas (such as lesions in medical images), which may result in blurring or information loss after compression.

[0058] 2. Blocking and redundancy: Misalignment between block boundaries and semantic content can easily lead to discontinuous blocking artifacts at block boundaries. Furthermore, relatively high bitrates may still be used for compression in non-critical areas, resulting in storage and bandwidth redundancy.

[0059] Secondly, global deep learning compression methods (such as end-to-end models based on CNN / Transformer, such as the Ballé 2018 architecture) have improved overall compression efficiency, but:

[0060] 1. Single compression ratio limitation: A unified compression strategy is used for the entire image, and the compression ratio cannot be dynamically adjusted based on the importance of the region. This results in significant quality degradation in key areas at high compression ratios, or waste of storage in non-critical areas at low compression ratios.

[0061] 2. Hardware performance bottleneck: When processing high-resolution images (such as 4K images), all image features need to be loaded at once, resulting in extremely high video memory usage (e.g., over 12GB), making it difficult to deploy efficiently on ordinary GPUs.

[0062] 3. Low serial processing efficiency: The decoding process of some models (especially those that rely on autoregressive context) cannot be parallelized, resulting in a significant increase in decoding latency and making it unsuitable for real-time applications.

[0063] Secondly, hybrid compression frameworks (such as ROI detection + traditional encoding such as JPEG / HEVC) attempt to combine the perception capabilities of deep learning with the maturity of traditional encoding, but they have the following drawbacks:

[0064] 1. Inefficient separate architecture: The ROI detection network and traditional encoder usually run independently, and feature extraction and compression coding cannot be jointly optimized, resulting in increased overall encoding and decoding delay.

[0065] 2. Poor feature compatibility: Traditional encoders (such as HEVC) cannot effectively utilize the rich semantic features extracted by deep learning (for example, super-prior information is difficult to be directly compatible with traditional quantization tables), resulting in limited improvement in compression rate.

[0066] 3. Insufficient context modeling: Existing hybrid solutions often have fixed anchor point selection (or region division) strategies for simple and complex regions, which cannot fully adapt to content changes and may result in redundant bitstreams or poor processing of complex textures.

[0067] Summary of common defects of existing technologies:

[0068] 1. Compression strategies are disconnected from semantic understanding: Traditional methods rely on manual features, while global deep learning models lack fine-grained control capabilities, and both fail to achieve effective dynamic coupling between compression strategies and image semantic information.

[0069] 2. Fragmentation of the encoding and decoding process: Especially in hybrid frameworks, feature extraction and compression encoding processes are separated, lacking end-to-end integrated design and optimization, resulting in efficiency loss and performance bottlenecks.

[0070] 3. Unoptimized binary structure: The binary code stream storage structure of existing methods is mostly not specifically designed for the characteristics of block compression. It lacks a flexible layered verification mechanism and dynamic parameter embedding capabilities, resulting in the need to improve encoding and decoding efficiency and robustness.

[0071] Therefore, the industry urgently needs a better way to solve the above problems.

[0072] In view of this, the present invention proposes a deep learning-based image block compression and decompression method, which effectively overcomes many shortcomings of the existing technology and achieves significant improvements in compression performance, visual quality and application flexibility. The technical solution of the present invention is shown in the following embodiments:

[0073] Example

[0074] like Figure 1 As shown, the present invention discloses a method for image block compression and decompression based on deep learning, comprising the following steps:

[0075] S1. Generate an importance evaluation matrix for the input image through the spatial attention network and dynamically divide the image into blocks;

[0076] S2, classifying image blocks into three categories according to their importance values;

[0077] S3, perform differential compression on the three types of blocks;

[0078] S4, encapsulating the compressed data into a three-dimensional binary structure including a header layer, a data layer, and a redundancy layer;

[0079] S5. At the decoding end, the image is reconstructed through parallel decoding units, and an edge fusion network is used to eliminate splicing artifacts.

[0080] In one feasible implementation, in step S1:

[0081] The S1 step is as follows:

[0082] S1.1. Process the input image using a feature extraction network based on the backbone network and integrated with the attention module, and output a feature map.

[0083] S1.2. Generate a block suggestion heatmap based on the feature map through the block decision head;

[0084] S1.3. Based on the feature map, the semantic importance value of each potential block position is calculated through multi-task learning to form an importance evaluation matrix, and the image blocks are dynamically divided according to the matrix.

[0085] In a specific embodiment, the backbone network adopts an improved ResNet50 network. A spatial attention module (SAM) is inserted between Stage 3 and Stage 4 of ResNet50 to enhance the capture of spatial context information. The feature extraction layer uses the first four stages of ResNet50 (removing the last global average pooling layer and the fully connected classification layer). The attention fusion module cascades the channel attention (SE Block) and the coordinate attention (CoordinateAttention) to simultaneously focus on the inter-channel dependency and precise spatial location information, further improving the richness and pertinence of the feature expression. The block decision head consists of one or more 3×3 deformable convolution layers followed by a Sigmoid activation function. Deformable convolution can adaptively adjust the sampling position according to the image content, so as to better adapt to target areas of different shapes and scales. The output is a block recommendation heat map, whose resolution is usually in a certain ratio to the original image (such as 1:16).

[0086] In one feasible implementation, the calculation formula of the importance evaluation matrix is:

[0087]

[0088] where σ is the Sigmoid function, GAP represents global average pooling, Fi is the feature map of the i-th layer, m is the scaling factor, whose initial value is 1.0 and is optimized through backpropagation.

[0089] In the above method, through multi-task learning, the network not only outputs block proposals, but also generates a quantization importance value S(x, y) within the range of [0, 1] for each potential pixel position (x, y) (or the grid cell corresponding to the heatmap resolution).

[0090] In an implementable embodiment, in step S2, the block includes a key block, a transition block, and a background block.

[0091] In an implementable embodiment, the key block is compressed using a residual dense block U-Net architecture, the transition block is compressed using a deformable convolution compression network, and the background block is compressed using quantization-aware MobileNetV3.

[0092] In an implementable embodiment, the key block adopts a U-Net architecture with 5-level residual dense blocks and is based on CABAC entropy coding.

[0093] Specifically, the processing of the key block (S(x, y)>0.8) is as follows:

[0094] Network model: Adopt a residual dense block U-Net architecture. Specifically, it is a U-Net architecture, and both the encoder and decoder contain 5-level residual dense blocks RDB. Each RDB contains 4 residual groups. The number of network channels gradually expands from 64 at the input end (encoder) to 512 at the deepest layer, and then symmetrically shrinks gradually (decoder). Entropy coding scheme: Adopt context-adaptive arithmetic coding (CABAC). Its context model can dynamically construct using the statistical characteristics of the importance value S(x, y) of adjacent encoded blocks or its quantization coefficients to improve coding efficiency.

[0095] In an implementable embodiment, the transition block adopts a compression network with 3 layers of deformable convolution and is dynamically adjusted through the quantization step size.

[0096] Specifically, the processing of the transition block (0.4 < S(x, y) ≤ 0.8) is as follows:

[0097] The network model uses a deformable convolutional compression network, primarily composed of three stacked deformable convolutional layers (DCNv2). This network supports dynamic adjustment of the receptive field to accommodate content of varying shapes, and the output channel compression ratio can be set to, for example, 1 / 4. Quantization control: A differentiable quantization layer (DSQ) is used. The quantization step size Δ is dynamically adjusted based on the importance value S(x,y) of the block, for example: Δ = 0.1 + 0.9 * S(x,y).

[0098] In one feasible implementation, the background block adopts the MobileNetv3 architecture embedded with an asymmetric quantization module and combines RLE and Huffman coding.

[0099] Specifically, the background block processing (S(x,y)≤0.4) steps are as follows:

[0100] The network model uses the quantization-aware MobileNetV3. For example, a miniaturized version of MobileNetV3 is used as the underlying architecture, and an Asymmetric Quantization Module (AQM) is embedded, supporting aggressive 8-bit fixed-point quantization (weights and activations). Entropy coding optimization combines run-length encoding (RLE) to compress consecutive zero values ​​(common after quantization), and then applies Huffman coding to non-zero coefficients and run lengths.

[0101] In the present invention, the three-dimensional binary structure of the bitstream organization is shown in the following table:

[0102]

[0103] The present invention also includes a decoding and reconstruction layer, which is mainly divided into a parallel decoding unit and an edge fusion network, as follows:

[0104] 1. Parallel decoding unit: Deploys GPU thread groups, each thread processes a 32×32 pixel block, and implements asynchronous execution through CUDA streams.

[0105] 2. Edge fusion network: uses a bidirectional gated recurrent unit (Bi-GRU) to smooth features along block boundaries and eliminate splicing artifacts.

[0106] The present invention also includes a quality enhancement module, which is as follows:

[0107] 1. Generator: ESRGAN improved architecture, embedding importance weight map as conditional input

[0108] 2. Discriminator: Multi-scale PatchGAN, focusing on supervising edge details of highly compressed blocks

[0109] The performance of the proposed method is evaluated by comparing it with existing mainstream compression methods on multiple standard test sets. The main evaluation indicators include peak signal-to-noise ratio (PSNR, unit: dB, higher is better), structural similarity (SSIM, range: 0-1, closer to 1 is better), and compression rate (%). Figure 2 As shown in the figure, the deep learning-based image block compression and decompression method proposed in this paper can achieve significantly better PSNR and SSIM values ​​at similar or higher compression rates than traditional JPEG, WebP, and a representative global deep learning compression method (Ballé 2018) for various types of image data (including professional medical images, satellite images, and general natural scene images). This demonstrates that the present invention has significant advantages in both compression efficiency and the objective quality of reconstructed images. It can effectively preserve image details and suppress compression artifacts, and is particularly outstanding in protecting information in key areas.

[0110] Another aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0111] The advantages of the present invention are as follows:

[0112] 1. Significantly improve compression performance and visual quality: Through semantically aware dynamic block segmentation and a multi-scale, differentiated compression strategy, this method can allocate the bit budget more intelligently and finely. As a result, at the same target bitrate, the reconstructed image can achieve higher objective evaluation metrics (such as PSNR and SSIM). At the same time, subjective visual quality is also improved, image details (especially those in important areas) are more fully preserved, and compression artifacts (such as blocking effects and blurring) are effectively suppressed.

[0113] 2. Enhanced semantic understanding and content adaptability: The deep learning model introduced in this paper can automatically perceive the semantic importance of different image regions, allowing the compression strategy to be dynamically and tightly coupled with the image content. Compared with the fixed compression strategies of traditional methods or the single compression strategies of early global deep learning models, this has greater internal content adaptability and flexibility.

[0114] 3. Optimized Codec Architecture and Efficiency: The end-to-end deep learning model avoids the separation and optimization difficulties inherent in hybrid frameworks between feature extraction and compression encoding modules. The block processing mechanism, combined with a parallel decoding architecture on the decoder side (e.g., utilizing CUDA for parallel processing on the GPU), significantly improves codec processing speed. A specially designed optimized bitstream structure also facilitates fast parsing and error checking, improving overall efficiency.

[0115] 4. Flexibility and Scalability: The block-based processing approach reduces the extremely high hardware requirements (especially video memory) for processing high-resolution images all at once, making this method easier to deploy and apply on hardware platforms with varying performance. The model parameters, block-based strategies, and importance thresholds can all be adjusted to specific application scenarios and requirements, providing excellent flexibility and scalability.

[0116] 5. Compatibility Considerations and Robustness Enhancement: By retaining some JPEG standard marker segments in the bitstream header, this method ensures a certain degree of basic metadata compatibility with legacy decoders. Furthermore, the inclusion of a checksum mechanism and optional redundancy layers (such as Reed-Solomon codes) in the bitstream structure enhances the robustness of the compressed data during storage and transmission.

Claims

1. A deep learning-based image block compression and decompression method, characterized in that: The following steps are involved: S1. Generate an importance evaluation matrix for the input image through the spatial attention network and dynamically divide the image into blocks; S2, classifying image blocks into three categories according to their importance values; S3, perform differential compression on the three types of blocks; S4, encapsulating the compressed data into a three-dimensional binary structure including a header layer, a data layer, and a redundancy layer; S5. At the decoding end, the image is reconstructed through parallel decoding units, and an edge fusion network is used to eliminate splicing artifacts.

2. The image block compression and decompression method based on deep learning according to claim 1, characterized in that: The S1 step is specifically as follows: S1.

1. Process the input image using a feature extraction network based on the backbone network and integrated with the attention module, and output a feature map. S1.

2. Based on the feature map, generate a block suggestion heat map through a block decision head; S1.

3. Based on the feature map, the semantic importance value of each potential block position is calculated through multi-task learning to form an importance evaluation matrix, and the image blocks are dynamically divided according to the matrix.

3. The image block compression and decompression method based on deep learning according to claim 2, characterized in that: The calculation formula of the importance assessment matrix is: Where σ is the Sigmoid function, GAP represents global average pooling, Fi is the feature map of the i-th layer, and m is the scaling factor, which is initially 1.0 and is optimized by backpropagation.

4. The image block compression and decompression method based on deep learning according to claim 1, characterized in that: In the step S2, the blocks include key blocks, transition blocks and background blocks.

5. The image block compression and decompression method based on deep learning according to claim 4, characterized in that: The key blocks are compressed using the residual dense block U-Net architecture, the transition blocks are compressed using the deformable convolutional compression network, and the background blocks are compressed using the quantization-aware MobileNetV3.

6. The image block compression and decompression method based on deep learning according to claim 5, characterized in that: The key blocks adopt a U-Net architecture with a 5-level residual dense block and are based on CABAC entropy coding.

7. The image block compression and decompression method based on deep learning according to claim 5, characterized in that: The transition block adopts a compression network with 3 layers of deformable convolution, which is dynamically adjusted by quantization step size.

8. The image block compression and decompression method based on deep learning according to claim 5, characterized in that: The background block adopts the MobileNetv3 architecture embedded with an asymmetric quantization module and combines RLE and Huffman coding.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images