An Adaptive Image Compression Method and System Based on Multi-Source Information Fusion

Abstract

This invention discloses an adaptive image compression method and system based on multi-source information fusion, relating to the field of image processing technology. The method includes: acquiring an image to be compressed and dividing it into multiple image blocks; extracting multi-dimensional features of spatial complexity, frequency domain complexity, and semantic importance for each image block, calculating a priority score, and generating a priority mask map; simultaneously, receiving channel state information fed back from the decoding end and generating a global compression control command; dynamically selecting corresponding compression paths for image blocks of different priorities according to the priority mask map and the global compression control command, wherein the compression paths include a high-fidelity path based on graph neural networks, a balanced path based on hybrid Transformer-CNN, and a high-compression path based on lightweight CNN; and finally encapsulating the compressed data of each image block into a bitstream according to priority from high to low. This invention can adaptively adjust the compression strategy according to the image content value and channel state, significantly improving the fidelity of key data areas and overall compression efficiency in dynamic and harsh channel environments such as deep space communication.

Description

Technical Field

[0001] This invention belongs to the field of digital image compression coding technology, specifically relating to an adaptive image compression method and system that can dynamically adjust the compression strategy according to the image content value and channel state in complex transmission environments. Technical Background Existing image compression technologies (such as JPEG, HEVC, and early deep learning methods) have two major shortcomings: one is that they treat all image content equally, and the other is that their compression strategies are static.

[0002] Traditional standards (JPEG / HEVC) use fixed transform and quantization matrices, which cannot dynamically adjust according to local image characteristics (edges, textures, flat areas). This results in critical details (such as crater edges and mineral zones) being treated the same as redundant background, leading to the loss of high-value information or wasted bandwidth. While early deep learning models improved performance, the limited receptive field of CNN structures made it difficult to capture global dependencies, and their local inductive bias prevented them from adaptively allocating bitrates based on global semantic information. Essentially, they still treated feature locations "equally," lacking the ability to distinguish between different levels of information redundancy. Although recent Transformer-based models have made breakthroughs in global modeling, their computational complexity is high, and their compression strategies remain "static."

[0003] In terms of compression strategies, existing solutions are mostly static open-loop systems, where the quantization step size and entropy coding method are fixed before encoding. This makes them unable to adapt to dynamic environmental changes such as channel bandwidth and bit error rate, which can easily lead to the loss of critical data or low transmission efficiency, especially in scenarios such as deep space communication. Summary of the Invention

[0004] This invention aims to solve the problem that existing image compression technologies lack joint perception of image content value and channel state, resulting in the loss of key information or low compression efficiency.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An adaptive image compression method and system based on multi-source information fusion includes the following steps: Step S1: Image Preprocessing and Multi-Dimensional Value Assessment Module S1.1: Obtain the original image I to be compressed and divide it into N×N non-overlapping image blocks {B1,B2,...,B...} n The preferred value is N=32.

[0006] S1.2: For each image patch Bᵢ, perform the following three feature extractions to generate a multidimensional value vector V(Bᵢ): S1.2.1: Space Complexity Evaluation: Calculate the standard deviation σ of pixel values ​​within an image block and the average gradient magnitude Ğ between adjacent pixels.

[0007] Space complexity score , where ω1 and ω2 are weighting coefficients.

[0008] S1.2.2: Frequency domain complexity evaluation: Perform a discrete cosine transform (DCT) on the image block Bᵢ to obtain the transform coefficient matrix F.

[0009] Based on the position of the DCT coefficients in the frequency domain, they are divided into three regions: low frequency (LF), mid frequency (MF), and high frequency (HF).

[0010] Calculate the ratio of high-frequency coefficient energy to total energy. .

[0011] Frequency domain complexity score .

[0012] S1.2.3: Semantic Importance Evaluation: A lightweight semantic segmentation network (e.g., a pruned model based on MobileNet-V3) is deployed at the encoder. An image patch Bᵢ is input into this network, and the network outputs the probability that it belongs to a "high importance category" (e.g., face, text, vehicle, mineral feature band). Semantic importance score .

[0013] S1.3: Calculate the overall priority score Pᵢ for each image block Bᵢ using a weighted fusion model. Wherein, α, β, and γ are preset fusion weights, and satisfy α+β+γ=1. Preferably, in general image compression, α=0.2, β=0.3, and γ=0.5; in hyperspectral remote sensing image compression, the weight of γ is increased.

[0014] S1.4: Based on two preset dynamic thresholds and The image patches are divided into three priority levels: like Marked as a high-priority block (Level-1); like This is marked as a medium priority block (Level-2). like It is marked as a low-priority block (Level-3).

[0015] Finally, a priority mask map (M) corresponding to the original image is generated.

[0016] Step S2: Link Status Awareness and Command Generation Module S2.1: Establish an out-of-band control link between the encoder and decoder using the BP or QUIC protocol of the extended Deep Space Delay Tolerant Network (DTN).

[0017] S2.2: The encoding end receives the channel status report from the decoding end in real time through this control link. This report includes at least: Instantaneous available bandwidth BW (bps); Current channel bit error rate (BER); Round-trip time (RTT) in milliseconds.

[0018] S2.3: The encoder determines the channel state by either querying a pre-defined "state-policy mapping table" or executing a lightweight decision function. Generate global compression control commands .

[0019] For example: If BW < 100kbps or BER > 1e-4, the instruction is "High Compression Ratio Mode" and "Enable Robust Entropy Coding".

[0020] If BW > 1 Mbps and BER < 1e-6, the instruction is "High Quality Mode" and "Enable High-Fidelity Arithmetic Coding".

[0021] Step S3: Dynamic Adaptive Compression Engine This engine includes a modular encoder pool containing multiple pluggable feature extraction network modules, quantizers, and entropy encoders.

[0022] S3.1: Strategy Loading and Initialization: The engine reads the strategy generated in step S2. Initialize global compression parameters, such as the basic quantization step size. Entropy coding mode selector, etc.

[0023] S3.2: Priority-based adaptive coding: The engine traverses the priority mask map (M) and processes each image block Bᵢ in descending order of priority.

[0024] For Level-1 (high priority) blocks: Objective: To maximize information fidelity.

[0025] Feature extraction: The first encoder (Encoder-HQ) is invoked. This encoder is based on a graph neural network (GNN) architecture. It dynamically constructs a graph structure of different degrees for each pixel node based on the local information redundancy of pixels within the block (e.g., measured by the "Scaled Residual Rating" (SRS) described in Patent 8) and performs adaptive information aggregation.

[0026] Quantization: Fine-grained quantization is employed. Quantization step size. ,in ≤1, to ensure minimal quantization loss.

[0027] Entropy encoding: Arithmetic encoding is used by default. However, when global instructions... When a "high error rate" is indicated, the system dynamically switches to run-length encoding with error protection capabilities.

[0028] For Level-2 (medium priority) blocks: Objective: To balance compression ratio and reconstruction quality.

[0029] Feature extraction: The second encoder (Encoder-Balanced) is invoked. This encoder is based on a hybrid Transformer and CNN architecture (as described in Patent 4 or 7), capturing details through local window self-attention and capturing context through global pooling or downsampling.

[0030] Quantization: Standard quantization is used. Quantization step size. .

[0031] Entropy coding: Arithmetic coding is used.

[0032] For Level-3 (low priority) blocks: Objective: Maximize compression ratio.

[0033] Feature extraction: The third encoder (Encoder-HC) is invoked. This encoder is a lightweight convolutional neural network that focuses on fast dimensionality reduction and redundancy removal.

[0034] Quantization: Coarse quantization is used. Quantization step size. Where θ3 >> 1. In extremely poor channel conditions, it can be determined according to... The instruction skips the encoding and transmission of some or all Level-3 blocks.

[0035] Entropy coding: Huffman coding is used.

[0036] S3.3: Bitstream encapsulation: Each compressed data block, its corresponding priority label (L1 / L2 / L3), the encoder ID used, the quantization step index, and the compressed version of the priority mask (M) are encapsulated together in priority order to form the final image compressed bitstream.

[0037] Step S4: Decoding and Reconstruction S4.1: The decoding end receives the bitstream and first parses out the global parameters and priority mask (M).

[0038] S4.2: Based on the header information of each data block, call the decoder (Decoder-HQ, Decoder-Balanced, Decoder-HC) that is symmetrical to the encoder to perform decoding.

[0039] S4.3: Reconstruct the final image based on the decoded transform coefficients and priority mask. . Attached Figure Description

[0040] Figure 1 The overall flowchart of the adaptive image compression coding method based on multi-source information fusion provided in the embodiment of the present invention is shown; Figure 2 This invention illustrates the structure of the multi-dimensional value assessment module in step S1 of this embodiment. Figure 3 This diagram illustrates the modular architecture of the dynamic adaptive compression engine in step S3 of this embodiment of the invention. Figure 4 The diagram shows the structure of the graph neural network encoder used in the high-fidelity path in this embodiment of the invention. Figure 5 The diagram shows the hybrid Transformer-CNN encoder structure used in the balanced path in this embodiment of the invention. Figure 6 A schematic diagram of the data structure of the image compression bitstream generated in an embodiment of the present invention is shown. Detailed Implementation

[0041] This embodiment provides a dynamic adaptive image compression coding method based on content-channel joint awareness. For example... Figure 1 As shown, the method includes the following steps: image preprocessing and multi-dimensional value assessment (S1), link state awareness and instruction generation (S2), dynamic adaptive compression coding (S3), and stream encapsulation and transmission (S4). The following section combines... Figures 2-3 Each step is explained in detail.

[0042] Step S1: Image Preprocessing and Multidimensional Value Assessment S1.1: Image Blocking. Obtain the original image I to be compressed, with a resolution of W×H. Divide the image into N×N non-overlapping pixel blocks, where N is preferably 32 or 64. If the image boundary is less than N, mirror padding is performed. In this embodiment, N=32, resulting in a total of Image blocks {B1,B2,...,B} m}

[0043] S1.2: Multi-dimensional feature extraction. For each image patch Bᵢ, the value scores of the following three dimensions are calculated in parallel: (1) Space complexity score S s (Bᵢ): Calculate the standard deviation σ of pixel values ​​within the block and the average gradient magnitude Ğ.

[0044] , where μ is the average pixel value within the block.

[0045] The gradient is calculated using the Sobel operator.

[0046] Then normalization is performed: S s The value ranges from [0,1], with larger values ​​indicating more complex textures.

[0047] (2) Frequency domain complexity score : Perform a discrete cosine transform (DCT) on the image block Bᵢ to obtain a 32×32 transform coefficient matrix F.

[0048] The low-frequency region is defined as the 8×8 coefficient in the upper left corner, and the high-frequency region is defined as all other coefficients.

[0049] Calculate the proportion of high-frequency energy: .

[0050] The higher the score, the richer the high-frequency details.

[0051] (3) Semantic importance score : This embodiment deploys a lightweight semantic segmentation network, MobileNet-V3-SSDLite, which has been pre-trained on the COCO dataset and fine-tuned for target tasks such as faces, vehicles, text, and crater edges.

[0052] After scaling the image patch Bᵢ to 224×224, it is input into the network, and the output is the probability that the patch belongs to the "high importance category". .

[0053] Direct order .

[0054] S1.3: Priority Score Fusion. A weighted summation model is used to calculate the overall priority score Pᵢ. In this embodiment, for general natural images, α=0.25, β=0.25, and γ=0.5 are used; for deep space remote sensing images, α=0.2, β=0.2, and γ=0.6 are used to emphasize the importance of scientific objectives.

[0055] S1.4: Priority level classification. Two adaptive thresholds are preset. and The threshold setting can be dynamically adjusted according to the target compression ratio. The preferred solution is: , .

[0056] Division rules: Level-1 (high priority): Pᵢ≥0.7, representing areas with complex edges, textures, or containing important objects.

[0057] Level-2 (medium priority): 0.3≤Pᵢ<0.7, representing a region with general texture.

[0058] Level-3 (low priority): Pᵢ<0.3, representing a smooth background area.

[0059] Finally, a priority mask map M corresponding to the position of the original image patch is generated.

[0060] Step S2: Link Status Awareness and Command Generation S2.1: Establish out-of-band control link. A low-latency feedback channel is established between the encoder and decoder via the BP (Bundle Protocol) append block of the extended DTN protocol or the DATAGRAM frame of the QUIC protocol.

[0061] S2.2: Real-time Status Monitoring. The decoder continuously monitors the following channel parameters and sends a channel status report to the encoder every 100ms: Instantaneous bandwidth (BW): Calculated by measuring the amount of data successfully received per unit time.

[0062] Bit error rate (BER): Estimated by the proportion of CRC check failures in decoded frames.

[0063] Round-trip time (RTT): Measured by sending a timestamp request.

[0064] S2.3: Generate compression control commands. The encoding end, based on the received channel state, queries a preset "state-policy mapping table" (as shown in Table 1) to generate global control commands. .

[0065] Step S3: Dynamic Adaptive Compression Coding This step uses a modular encoder pool containing three optional encoding paths. For example... Figure 3 As shown, the engine dynamically selects a path for each image patch based on the priority mask map and global instructions.

[0066] S3.1: Initialize global parameters. According to... Set the basic quantization step size (Default value is 16) and entropy coding mode.

[0067] S3.2: Encoding block by block according to priority: Level-1 block processing (high-fidelity path): Feature extraction: A graph neural network encoder (Encoder-GNN) is invoked. Its structure is as follows: Figure 4 As shown.

[0068] First, the input block is mapped to a feature map using a 3×3 convolutional layer. Dimensions: 32×32×64.

[0069] Then, for each pixel node in the feature map, its "Scaling Residual Score" (SRS) is calculated. The larger the value, the lower the information redundancy.

[0070] Dynamically assign degree d to each node based on the SRS value: In other words, the more information there is, the more connections there are.

[0071] Within a local 8×8 window, candidate adjacent nodes are densely sampled, sorted by feature similarity, and the top d nodes are selected as adjacent nodes to construct a local graph structure.

[0072] Information is aggregated using a two-layer graph attention network (GAT) to update the features of each node. The graph aggregation operation is defined as follows: ,in This represents the attention coefficient.

[0073] Output feature map F1, size 32×32×128.

[0074] Quantization: Fine-grained quantization is employed. Quantization step size. .

[0075] Entropy coding: If the global instruction is MODE_HIGH_QUALITY, arithmetic coding is used; if it is MODE_ROBUST, run-length coding is switched to MODE_ROBUST.

[0076] Level-2 block processing (balanced path): Feature extraction: A hybrid Transformer-CNN encoder (Encoder-Hybrid) is invoked. Its structure is as follows: Figure 5 As shown.

[0077] First, downsampling is performed using a convolution with a stride of 2 to obtain a 16×16×128 feature map.

[0078] Then, four local-global feature aggregation blocks are stacked sequentially. Each block consists of: a 3×3 depthwise separable convolution (extracting local features), a 4×4 SwinTransformer block (extracting global features), and finally fused by channel concatenation and a 1×1 convolution.

[0079] Output feature map F2, size 16×16×256.

[0080] Quantification: Standardized quantification. .

[0081] Entropy coding: Arithmetic coding is used.

[0082] Level-3 block processing (high compression ratio path): Feature extraction: A lightweight CNN encoder (Encoder-Light) is invoked. Its structure consists of three cascaded convolutional blocks, each containing a 3×3 convolution with a stride of 2 (downsampling) and a ReLU activation. The final output is a feature map. Dimensions: 4×4×512.

[0083] Quantification: Coarse quantification. Under the MODE_ROBUST command, further... Increased to 4 times.

[0084] Entropy coding: Huffman coding is used.

[0085] S3.3: Stream encapsulation and transmission. For example... Figure 6 As shown, each encoded data block is encapsulated according to priority order (Level-1 first, then Level-2, and finally Level-3). The header information of each data block includes: priority tag (2 bits), encoder ID (2 bits), quantization step size index (4 bits), and the position index of the block in the priority mask map. After all blocks are concatenated, the compressed priority mask map is appended to the end of the bitstream to form the final image compressed bitstream.

[0086] Step S4: Decoding and Reconstruction After receiving the bitstream, the decoder first parses out the priority mask and global parameters. Then, based on the header information of each data block, it calls a symmetric decoder (Decoder-GNN, Decoder-Hybrid, Decoder-Light) to perform dequantization and feature reconstruction. Finally, based on the priority mask, the decoded image blocks are stitched together at their corresponding positions to obtain the reconstructed image I'.

Claims

1. An adaptive image compression encoding method based on multi-source information fusion, characterized in that, Includes the following steps: Step S1: Obtain the image to be compressed and divide the image to be compressed into multiple image blocks; Step S2: Perform multi-dimensional feature extraction on each image block, calculate the priority score of each image block based on the extracted multi-dimensional features, and divide the image block into at least two priority levels based on the priority score to generate a priority mask map; Step S3: Receive channel state information fed back from the decoding end, and generate a global compression control command based on the channel state information; Step S4: Based on the priority mask map and the global compression control command, dynamically select the corresponding compression path for each image block, wherein different priority levels correspond to different compression paths, and the compression path includes at least a combination of feature extraction network, quantization step size and entropy coding method; Step S5: According to the order of priority from high to low, the compressed data of each image block is encapsulated into a bitstream, and the priority mask map is also encapsulated to generate the final image compressed bitstream.

2. The adaptive image compression encoding method based on multi-source information fusion according to claim 1, characterized in that, The multi-dimensional feature extraction in step S2 includes: Calculate the spatial complexity score of the image patch, which is determined based on the standard deviation and average gradient magnitude of the pixel values ​​within the image patch; Calculate the frequency domain complexity score of the image patch, which is determined based on the proportion of high-frequency coefficient energy to the total energy after performing a discrete cosine transform on the image patch; The semantic importance score of the image patch is calculated by inputting the image patch into a pre-trained lightweight semantic segmentation network.

3. The adaptive image compression encoding method based on multi-source information fusion according to claim 2, characterized in that, The priority score is calculated using a weighted summation model: wherein S s is a spatial complexity score, is a frequency domain complexity score, is a semantic importance score, and a, b, g are preset weight coefficients satisfying .

4. The adaptive image compression encoding method based on multi-source information fusion according to claim 1, characterized in that, The step S2, which involves classifying priority levels based on priority scores, includes: a first threshold value and a second threshold value , and ; When a high priority level is assigned; when At that time, it is classified as a medium priority level; when When that happens, it is classified as a low priority level.

5. The adaptive image compression coding method based on multi-source information fusion according to claim 1, characterized in that, The channel state information in step S3 includes at least one of instantaneous bandwidth, bit error rate, and round-trip time; The global compression control command includes at least a compression mode indicator and an entropy coding mode indicator.

6. The adaptive image compression coding method based on multi-source information fusion according to claim 1, characterized in that, The compression paths corresponding to different priority levels in step S4 include: For high-priority image blocks, a first compression path is adopted, which includes: feature extraction using a graph neural network encoder, quantization using a first quantization step size, and entropy encoding using arithmetic coding or run-length coding. For medium-priority image blocks, a second compression path is used, which includes: feature extraction using a hybrid Transformer and convolutional neural network encoder, quantization using a second quantization step size, and entropy encoding using arithmetic coding. For low-priority image blocks, a third compression path is used, which includes: feature extraction using a lightweight convolutional neural network encoder, quantization using a third quantization step size, and entropy encoding using Huffman coding. Wherein, the first quantization step size is smaller than the second quantization step size, and the second quantization step size is smaller than the third quantization step size.

7. The adaptive image compression coding method based on multi-source information fusion according to claim 6, characterized in that, The graph neural network encoder is used for: Map the input image patches to the initial feature map; Calculate the scaling residual score for each pixel node in the initial feature map, where the scaling residual score is the difference between the initial feature map and the feature map after downsampling and then upsampling. The degree of each pixel node is dynamically allocated based on the scaling residual score, and the degree is inversely proportional to the information redundancy of the node. Candidate neighboring nodes are collected within a local window of each pixel node, and the number of neighboring nodes of the specified degree are selected according to feature similarity to construct a graph structure. The graph structure is aggregated using a graph attention network to update the feature representation of each pixel node.

8. The adaptive image compression coding method based on multi-source information fusion according to claim 6, characterized in that, The hybrid Transformer and convolutional neural network encoder includes: At least one downsampling convolutional layer; Multiple locally-global feature aggregation blocks are stacked sequentially. Each locally-global feature aggregation block includes: a depthwise separable convolutional layer for extracting local features, a SwinTransformer block for extracting global features, and a channel concatenation layer and a 1×1 convolutional layer for fusing local and global features.

9. The adaptive image compression coding method based on multi-source information fusion according to claim 1, characterized in that, The specific format for encapsulating the bitstream in step S5 includes: The bitstream header contains the image size, block size, and basic quantization step size parameters; The main body of the bitstream consists of compressed data of each image block arranged in descending order of priority. The compressed data of each image block includes a block header and a block body. The block header contains a priority label, encoder identifier, quantization step size index and position index, and the block body is the encoded binary data. The end of the bitstream contains compressed priority mask data.

10. The adaptive image compression coding method based on multi-source information fusion according to claim 1, characterized in that, Also includes: When the channel state information indicates that the bit error rate is higher than a preset threshold, the entropy coding method of high-priority image blocks is switched from arithmetic coding to run-length coding or Huffman coding.

Images