Image compression reconstruction method and system based on block modulation sequence flow compression

Abstract

The invention discloses an image compression reconstruction method and system based on block modulation sequence stream compression. The method comprises the following steps: acquiring an input image and a modulation mask; dividing the modulated image into non-overlapped sub-blocks, and extracting pixels at the same position from each sub-block to form a low-resolution image sequence flow; summing pixels at the same position in all the low-resolution sequence flow images according to index dimensions to generate compression measurement; converting reconstruction into an optimization problem according to a forward measurement model, and constructing an interpretable depth expansion mathematical model; the cascade design of a cross attention enhancement module, a multi-mode expansion Mama module and a channel attention module is combined in a deep denoising approximation operator, and a high-resolution image is efficiently reconstructed through a deep expansion network. The low-complexity forward coding model and the high-efficiency decoder design provided by the invention optimize the coding and decoding efficiency, maintain excellent reconstruction precision while greatly improving the compression ratio, and are suitable for scenes such as a mobile terminal monitoring platform with limited resources.

Description

Technical Field

[0001] This invention relates to the field of computational imaging and image compression, and in particular to an image compression and reconstruction method and system based on block modulation sequence stream compression. Background Technology

[0002] Infrared imaging technology has important applications in UAV reconnaissance, night vision security, aerospace remote sensing, and resource-constrained edge intelligent systems. However, such systems typically rely on high-resolution focal plane array detectors to sample the entire image pixel by pixel. This not only increases the manufacturing and deployment costs of the devices but also requires high data bandwidth and energy consumption during real-time transmission and storage, making efficient deployment difficult for platforms with limited computing power (such as UAVs and mobile terminals). Therefore, low-sampling imaging methods based on compressed sensing theory have attracted attention, aiming to improve transmission and storage bottlenecks by reducing the amount of sampled data.

[0003] For example, Chinese patent document CN111193925A discloses an image compressed sensing coding and normalization method based on block vector inner product; Chinese patent document CN103037212A discloses a low-complexity adaptive block compressed sensing image coding method based on visual perception.

[0004] While existing block compression coding schemes reduce computational load at the encoding end to some extent, the encoding process requires repeated use of each block, making its forward process difficult to implement in real infrared optical systems. Furthermore, current coding methods typically treat image grayscale values ​​as carriers of structural information, neglecting the fact that pixel values ​​in infrared images are essentially thermal radiation intensity, significantly coupled with the target object's temperature, material emissivity, and spectral distribution. Ignoring this physical property leads to the inability to effectively preserve the true radiation field distribution characteristics during the encoding stage, resulting in reconstruction degradation in infrared scenes with low signal-to-noise ratios, weak textures, and long-range targets.

[0005] On the other hand, existing block-based compression methods often disrupt the global continuity of images and are insufficient in modeling long-range correlations across block regions, making them prone to detail loss, block artifacts, and edge distortion during the decoding stage. Furthermore, most current decoding algorithms are designed for high-performance computing platforms, making real-time inference difficult to achieve on edge devices.

[0006] Therefore, there is an urgent need for a novel compressed imaging scheme that simultaneously considers physical feasibility, energy consistency, and edge deployment, so that infrared images can retain effective radiation information structure during the encoding stage and achieve low-complexity, high-fidelity reconstruction during the decoding stage. Summary of the Invention

[0007] This invention provides an image compression and reconstruction method and system based on block modulation sequence stream compression. By introducing a pixel-by-pixel scanning block aggregation mechanism at the encoding end and adopting an end-to-end trained iterative depth expansion recovery structure at the decoding end, high compression ratio encoding and high-fidelity reconstruction of images can be achieved.

[0008] An image compression and reconstruction method based on block modulation sequence stream compression includes the following steps: (1) Obtain the input image and a modulation mask with the same size as the image, and complete the modulation of the image by pixel-by-pixel multiplication; (2) During the encoding process, the modulated image is divided into non-overlapping sub-blocks according to the preset block size, and then the pixels at the same position are extracted from each sub-block to form a low-resolution image sequence stream; (3) Summate the pixels at the same position in all images in the low-resolution image sequence stream along the block index dimension, and concatenate them in row and column order to generate compressed measurement y; (4) During the decoding process, the high-resolution image to be recovered from the compressed measurement y is transformed into an optimization problem; (5) Iterate according to the GAP algorithm The above optimization problem is solved by constructing a deep unfolding mathematical model. In each iteration, the internal deep denoising approximation operator is solved by a cascaded design of cross-attention enhancement module, multi-mode unfolding Mamba module and channel attention module, and finally reconstructs low-resolution sequence flow sub-image. (6) Perform the inverse transformation operation of the encoding dual on the reconstructed low-resolution sequence sub-image, reassemble and fuse the sub-image information in the spatial dimension, restore the full-resolution image of the same scale as the input image, and complete the final decoding output.

[0009] In step (1), for the size of Image to be compressed Use a binary encoding board of the same size M modulation.

[0010] In step (2), according to The size divides the encoded image to be compressed into m There are 1 non-overlapping sub-blocks, and the number of blocks is 1. .

[0011] In step (3), extract sequentially m Pixels at the same position in each block (for example, extracting the first pixel of each block can yield a low-resolution sub-image of the original image, with a size of...) This can yield a sequence of low-resolution sub-images from the original high-resolution image, ranging from 1 to... m The sequence of sub-images is called a sequence stream. X The size of the sequence stream is Generate compression measurement The process is to sum the sequence stream, which can be expressed as: ; in, Represents the first in the sequence stream i A low-resolution sub-image, with a size of [number]. , This represents the final encoded compression measurement, the size of which is compared to the low-resolution sub-unit. Figure 1 The compression ratio is... .

[0012] In step (4), the compression measurement will be performed. Recovering a high-resolution image from a medium can be transformed into the following optimization problem: ; Where y represents the input compression measurement, X This represents a stream of low-resolution sub-image sequences to be reconstructed. H This indicates that the binary encoding board will be used. M The same low-resolution coded mask sequence is obtained by the same decomposition process as the image. Denotes the a priori regularization term. It is a balance parameter.

[0013] In step (5), according to the GAP iteration The following is a method for solving the optimization problem by constructing a deep unfolded mathematical model, including measurement consistency correction and prior constraint recovery, as detailed below: Measurement consistency correction: in the first In the current iteration, using the current... estimation of the next iteration The deviation between y and y is used to construct a direction correction term, which is then combined with the operator. The residuals are projected onto the image domain to obtain intermediate variables. The expression is: ; in, The learnable step size parameter is used to control the projection intensity, enabling the model to adaptively select the update magnitude at different stages; Prior constraint recovery: for intermediate variables Apply a learnable prior map to obtain a next-stage estimate that conforms to the statistical properties of the natural image. Its form is: ; in, Indicates the first A learnable deep denoising prior operator is implemented to map low-quality sequential image streams to high-fidelity sequential image stream representations. The learnable prior operator consists of a cascaded design of a cross-attention enhancement module, a multi-modal unfolded Mamba module, and a channel attention module. The input... First, the image is transferred from the image domain to the feature domain by the shallow feature extraction module, and then it participates in the calculation of the subsequent neural network module.

[0014] In step (5), the cross-attention enhancement module is implemented as follows: The deep features output from the previous stage are first concatenated with the initial features of the current stage along the channel dimension to obtain the cross-stage concatenated input features. ; Input features after cross-stage concatenation Global average pooling is performed to obtain channel-level statistical descriptions, and attention weights are generated through a layer-by-layer mapping function; the attention weights are then compared with the enhanced output features. Represented in the following form: ; in, Indicates global average pooling; and Represents a learnable linear mapping; It is a non-linear activation function; For the Sigmoid function; This is a channel-by-channel multiplication operation.

[0015] In step (5), the Mamba module with multi-mode expansion is implemented as follows: For the input features of the Mamba module with multi-modal expansion, it is expanded in three modes: spatial horizontal expansion, spatial vertical expansion, and channel-dimensional expansion, resulting in sequences. The recursive calculation process within each mode can be described as follows: ; After discretization, its evolutionary update form is as follows: ; The corresponding features are output from the hidden state: ; in, It is a state vector; These are learnable parameters; To address the discrepancy in contribution among different modes, a set of learnable weighted parameters is further introduced. This is used to dynamically adjust the fusion priority of different modes. The fused output is represented as: ; in, express The three unfolding modes of the data.

[0016] In step (5), the channel attention module is implemented as follows: The input features are first rearranged according to the channel direction, flattening the two-dimensional space into a one-dimensional sequence representation, and then the query is obtained through linear mapping. Q ,key K AND value V Three tensors; the input feature tensor is represented as Then its linear transformation is expressed as: ; ; ; in, These are the parameters for network learning.

[0017] Then, the corresponding channel correlation matrix is ​​weighted to obtain the attention map, and based on this, the feature selection and fusion are completed, resulting in enhanced channel features. Specifically, it is expressed as: ; ; in, To normalize the scale and stabilize the attention weight distribution, a residual fusion method is finally used to integrate the enhanced channel features. Compared with the original input features Element-wise superposition is performed to form a stable enhanced output. To achieve non-local expression enhancement at the semantic level: .

[0018] An image compression and reconstruction system based on block modulation sequence stream compression includes a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they implement the above-mentioned image compression and reconstruction method.

[0019] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the above-described image compression and reconstruction method.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention reduces the computational complexity of high-resolution image compression by decomposing a high-resolution image into a low-resolution sub-image sequence stream for encoding and compression, while fully preserving the low-frequency information and long-distance pixel correlation of the original image, providing sufficient physical constraints for subsequent reconstruction algorithms.

[0021] 2. This invention optimizes the measurement consistency correction and prior recovery of images by introducing an expansion algorithm that combines generalized alternating projection (GAP) with a priori operators of deep neural networks. The cascaded design of the cross-attention enhancement module, the multi-mode expansion Mamba module and the channel attention mechanism in the deep denoiser further improves the accuracy of sequence stream reconstruction and effectively reduces the computational complexity for large-size image reconstruction.

[0022] 3. This invention provides a new approach for the design of low-complexity, high-resolution compressed sensing coded imaging systems. The forward coding process of low-resolution sequence stream compression is equivalent to a weighted summation of the contents of each non-overlapping sub-block of the image. Therefore, in practical physical system design, it is possible to consider introducing metasurface phase coding to encode different regions of the target so that they can be captured by different specific regions of the detector, thereby achieving highly integrated, low-cost, portable parallel compressed sensing imaging. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in 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.

[0024] Figure 1 This is a schematic diagram of the structure for generating compressed measurement values ​​in this invention.

[0025] Figure 2 This is a schematic diagram of the structure for solving the depth expansion mathematical model in this invention.

[0026] Figure 3 This is a schematic diagram of the cross-attention enhancement module in this invention.

[0027] Figure 4 This is a schematic diagram of the structure of the Mamba module with multi-mode expansion in this invention.

[0028] Figure 5 This is a schematic diagram of the channel attention module in this invention.

[0029] Figure 6 This is a schematic diagram of the physical system design of the present invention.

[0030] Figure 7This is a schematic diagram of the image encoding and decoding in an example scenario application of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0033] This invention achieves low-cost encoding of the entire image by dividing the input image into low-resolution block patches and using an embedded scanning mapping method to achieve pixel-by-pixel cumulative compression within each block. Unlike previous methods that involved "block division and summation," this invention employs pixel-by-pixel position-based scanning aggregation, which reduces encoding complexity while preserving more recoverable structural information. At the decoding end, this invention uses a deep unfolded network architecture, jointly modeling spatial and channel-dimensional correlations through a multimodal unfolded sequence modeling module and a channel-related self-attention module, achieving high-fidelity reconstruction of low-resolution block sequences.

[0034] like Figures 1-2 As shown, an image compression and reconstruction method based on block modulation sequence stream compression includes the following steps: S1. Obtain the input image and a modulation mask of the same size. Modulation is achieved through pixel-by-pixel multiplication, resulting in... ,in This represents pixel-level multiplication. This modulation operation is used to improve the controllable representation of spatial distribution information during the encoding stage.

[0035] S2, the image to be modulated Divided according to preset block size m The size is For each non-overlapping sub-block, extract the same coordinate position. The pixels form a low-resolution sub-image sequence stream. X Size is , It provides position-aligned statistical units for pixel-by-pixel scan aggregation.

[0036] S3. Sum the low-resolution image sequence stream to generate a compressed measurement. , ; in, Represents the first in a low-resolution image sequence stream i A low-resolution sub-image.

[0037] like Figure 1 As shown, firstly, the input image is modulated pixel-by-pixel with a random binary mask of the same size to obtain the modulated image. Then, the modulated image is divided into non-overlapping sub-blocks according to a preset block size, each sub-block corresponding to a local spatial region in the original image. Based on this, instead of directly summing the entire block, this invention extracts pixels with the same relative coordinate positions in all sub-blocks, forming a set of low-resolution image sequence streams. By continuing to perform pixel-by-pixel accumulation operations across the block dimension on these low-resolution image sequence streams, the final compressed measurement image is obtained. It can be seen that this scheme essentially achieves "intra-block pixel-by-pixel scanning aggregation." This mapping method avoids the superposition and aliasing of distant pixels, converting a high-resolution two-dimensional image into a low-dimensional recoverable measurement, providing a structurally consistent compressed representation for the subsequent decoding stage.

[0038] S4. The inverse problem of recovering a high-resolution image from compressed measurements is transformed into solving the following optimization problem: ; Where y represents the input compression measurement, X This represents the low-resolution sub-image sequence stream to be reconstructed. H This indicates that the binary encoding board will be used. M The resulting low-resolution coded mask sequence, obtained through the same decomposition process as the image, has a size of [size missing]. , Denotes the a priori regularization term. It is a balance parameter.

[0039] S5. Iterate according to the GAP algorithm. The deep unfolding mathematical model is constructed to solve S4. In each iteration, the deep denoising approximation operator is solved by a cascaded design of the Cross-stage Channel Attention Block (CCAB), the Multimode Unrolling Mamba block (MUMB), and the Channel-wise Attention Block (CSAB).

[0040] S5.1 In the decoding stage of this invention, a learnable deep iterative expansion solution mechanism based on GAP (Generalized Alternating Projection) is introduced to progressively approximate and recover the compressed measurement. This iterative process consists of two parts: "measurement consistency correction" and "prior constraint recovery," which are repeatedly executed in each iteration to achieve approximate reconstruction of the target image.

[0041] S5.1.1 Measurement consistency correction: In the first... k In the current iteration, using the current... k -1 iterations of estimation The deviation between y and y is used to construct a direction correction term, which is then combined with the operator. The residuals are projected onto the image domain to obtain intermediate variables. The expression is: ; in, This is a learnable step size parameter used to control the projection intensity, enabling the model to adaptively select the update magnitude at different stages.

[0042] S5.1.2, Prior Constraint Recovery: Subsequently, for intermediate variables... Apply a learnable prior map to obtain a next-stage estimate that conforms to the statistical properties of the natural image. Its form is: ; in, Indicates the first The learnable prior operators for each stage are used to perform operations such as noise suppression, structure compensation, and texture restoration in the deep feature domain, thereby realizing the mapping from low-dimensional, low-quality sequential stream images to high-fidelity image representations. The iterative solution method is as follows: Figure 2 As shown, the deep denoising approximation operator is implemented using a cascaded design of three main modules: the Cross-stage Channel Attention Block (CCAB), the Multimode Unrolling Mamba block (MUMB), and the Channel-wise Attention Block (CSAB).

[0043] S5.1.2.1 Implementation of the Cross-Attention Enhancement Module (CCAB): This invention sets up a cross-stage channel attention module in the deep unfolding structure of the decoder to achieve adaptive fusion of deep features from the previous stage and shallow features from the current stage. This module performs channel-dimensional statistics and selective enhancement on the cross-stage input features, enabling historical information to participate in the current stage inference in a weighted manner, thereby achieving cross-level and cross-stage information backflow. In actual implementation, the deep features from the previous stage output and the initial features of the current stage are first concatenated along the channel dimension to obtain a fused input feature tensor. Subsequently, this invention performs global average pooling on the concatenated features to obtain channel-level statistical descriptions and generates attention weights through a layer-by-layer mapping function. These weights are used to weight the channel dimension element-wise, amplifying important channels and suppressing irrelevant or redundant channels, thereby enhancing the discriminative power of feature representation.

[0044] Specifically, such as Figure 3 As shown, the deep features output from the previous stage are first concatenated with the initial features of the current stage along the channel dimension to obtain a fused input feature tensor. Subsequently, this invention performs global average pooling on the concatenated features to obtain channel-level statistical descriptions and generates attention weights through a layer-by-layer mapping function. These weights are used to perform element-wise weighting along the channel dimension, amplifying important channels and suppressing irrelevant or redundant channels, thereby enhancing the discriminative power of the feature representation. Let the input features after cross-stage concatenation be... Then, the output features generated and enhanced by attention weights can be represented in the following form: ; in, Indicates global average pooling; and Represents a learnable linear mapping; It is a non-linear activation function; For the Sigmoid function; This is a channel-by-channel multiplication operation. Through the above mechanism, the cross-stage channel attention module of the present invention can explicitly model the importance of different channel features in each stage, so that deep structural information continues to play a role in the iteration process, improve the effectiveness of feature transmission and the stability of overall reconstruction, and effectively support the subsequent multimodal global modeling process.

[0045] S5.1.2.2 Implementation of the Multimodal Unfolded Mamba Module (MUMB): To solve the problem of limited local spatial correlation in a single unfolded mode, this invention sets up a multimodal unfolded sequence modeling module (MUMB) in the deep unfolded decoding structure to serialize local spatial features in multiple dimensions, thereby capturing long-range dependency information in the low-dimensional pseudo-sequence domain.

[0046] like Figure 4 As shown, this module first treats the input 3D feature map as a 2D tensor in multiple modes, and expands it into equivalent sequences with different arrangements along the channel dimension, horizontal dimension, and vertical dimension, thereby forming a multimodal representation. Based on this, the present invention introduces a linear state-space model for temporal recursive calculation for each expansion mode, realizing linear complexity modeling within the global receptive domain.

[0047] Input characteristics of multimodal unfolded Mamba modules The sequences were obtained by expanding them in three different modes. The recursive calculation process within each mode can be described as follows: ; After discretization, its evolutionary update form is as follows: ; The corresponding features are output from the hidden state: ; in, It is a state vector; These are learnable parameters; To address the issue of differing contributions among multimodal expansions, this invention further introduces a set of learnable weighting parameters. This is used to dynamically adjust the fusion priority of different modes, enabling the model to adaptively preserve more significant expansion paths during the inference phase. The fused output can be represented as: ; in, This invention represents three expansion modes of y. Through the aforementioned multimodal expansion and linear state-space recursion mechanism, it can capture a wide range of spatial dependencies without increasing explicit global attention overhead, while maintaining low inference complexity. This module effectively compensates for the lack of local context caused by traditional single expansion methods, providing a stable global structural basis for subsequent channel-dimensional augmentation calculations.

[0048] S5.1.2.3 Implementation of the Channel-Related Self-Attention Module (CSAB): To further enhance the semantic expressiveness of features, this invention sets up a Channel-Related Self-Attention Module (CSAB) after the MUMB module to model the global semantic dependencies between different channels. Unlike traditional spatial attention mechanisms, this module achieves cross-regional semantic reorganization at the channel dimension, significantly improving cross-regional structural expressiveness without increasing spatial resolution computational overhead.

[0049] like Figure 5As shown, in the implementation process, the input features are first rearranged according to the channel direction, flattening the two-dimensional space into a one-dimensional sequence representation, and then the query is obtained through linear mapping. Q ),key( K ) and value ( V Three tensors. If the input feature tensor is represented as Then its linear transformation can be expressed as: ; ; ; Then, the corresponding channel correlation matrix is ​​weighted to obtain the attention map, and feature selection and fusion are completed accordingly, which can be specifically expressed as: ; ; in, To normalize the scale and stabilize the distribution of attention weights, this invention employs a residual fusion method to element-wise superimpose the enhanced channel features with the original input features, forming a stable enhanced output and achieving non-local semantic expression enhancement. ; Through this self-attention mechanism, the present invention can complete the global reorganization of channel semantics without destroying spatial structural information, thereby improving the reconstruction quality of the decoding network for structurally sensitive regions (such as edges, textures, and repetitive patterns) and providing representational features with enhanced expressive power for the final reconstruction stage.

[0050] S6. The reconstruction result obtained after using the pre-trained iterative deep expansion algorithm. This refers to the sequence stream representation of low-resolution images. These sequence streams need to be restored to the full-resolution representation of the original images through an inverse mapping operation. This inverse mapping operation is a process of spatial restoration of block-level features, restoring the pixel positions of each low-resolution image in the sequence stream to the corresponding sub-blocks of the original high-resolution image, and stitching them back to the high-resolution structure of the original image. The restoration process is described by the following formula: ; in, This represents the restored high-resolution image. This represents the inverse mapping operator, which accurately stitches the low-resolution image sequence stream back into the original space by reversing the forward extraction process.

[0051] Figure 6A schematic diagram of the encoder physical system design of this invention is shown. The core design utilizes a metasurface to achieve phase encoding of high-resolution image blocks and mapping to specific regions of the detector. The phase distribution must ensure that the light intensity distribution formed in specific non-overlapping regions of the detector after modulation of the light field of each sub-block corresponds to the weighted sum of the encoded modulation. The metasurface can be monolithically integrated in front of the detector to form a thin, portable system. Metasurface materials can be selected from silicon dioxide, silicon, etc., and processing techniques include electron beam lithography or nanoimprint lithography. For specific wavelengths (such as visible light or near-infrared), the selection needs to be based on the application; broadband imaging may require dispersion compensation design. In the actual system, the system matrix, including the metasurface and detector responses, needs to be calibrated and measured to implement the backend decoding and reconstruction algorithm.

[0052] Figure 7 This paper demonstrates an example application scenario of the present invention, which involves compressing and capturing a target scene and then transmitting it to a terminal platform for reconstruction, thereby significantly reducing the bandwidth requirements for transmitting high-resolution images.

[0053] First, the compression coding system designed in this invention is deployed on the mobile terminal platform. After the target scene under test is compressed and coded into a sequence stream, the high-resolution image is compressed into a low-dimensional measurement and transmitted to the terminal platform. The decoder designed in this invention can recover the original high-resolution target scene from the low-dimensional measurement signal, realizing low-bandwidth storage and transmission of high-resolution target images.

[0054] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image compression reconstruction method based on block-wise modulation sequence stream compression, characterized by, Includes the following steps: (1) Obtain the input image and a modulation mask with the same size as the image, and complete the modulation of the image by pixel-by-pixel multiplication; (2) During the encoding process, the modulated image is divided into multiple non-overlapping sub-blocks according to the preset block size, and then the pixels at the same position are extracted from each sub-block to form a low-resolution image sequence stream; (3) Sum all the pixels in the same position along the block index dimension of all the images in the low-resolution image sequence stream, and stitch them in row-column order to generate a compressed measurement ; (4) During the decoding process, the high-resolution image to be recovered from the compressed measurement y is transformed into an optimization problem; (5) According to the GAP algorithm iteration Subsequently, a deep expansion mathematical model is constructed to solve the above optimization problem. In each iteration process, the internal deep denoising approximation operator is solved by cascading a cross-attention enhancement module, a multi-mode expansion Mamba module, and a channel attention module. Finally, a low-resolution sequence flow sub-image is reconstructed. (6) Perform the inverse transformation operation of the encoding dual on the reconstructed low-resolution sequence sub-image, reassemble and fuse the sub-image information in the spatial dimension, restore the full-resolution image of the same scale as the input image, and complete the final decoding output.

2. The image compression reconstruction method based on block modulation sequence stream compression according to claim 1, characterized in that, In step (4), recovering the high-resolution image from the compressed measurement y is transformed into solving an optimization problem, specifically: ； where y denotes the input compressed measurements, X denotes a sequence of low resolution sub-images to be reconstructed, H denotes a binary coded board M a sequence of equal-sized low resolution coded masks resulting from the same decomposition process as the image, denotes a priori regularizer, is a balancing parameter.

3. The image compression and reconstruction method based on block modulation sequence stream compression according to claim 1, characterized in that, In step (5), iteration is performed according to the GAP algorithm The above optimization problem is solved by constructing a deep expansion mathematical model, including measurement consistency correction and prior constraint recovery, as follows: Measurement consistency correction: in the first iteration, the deviation between the estimate of the current first iteration and y is used to construct a direction correction term, combined with the operator The residual is projected to the image domain to obtain an intermediate variable , expressed as: , the expression is: ​ ； in, The learnable step size parameter is used to control the projection intensity, enabling the model to adaptively select the update magnitude at different stages; Prior constraint recovery: for intermediate variables Apply a learnable prior map to obtain a next-stage estimate that conforms to the statistical properties of the natural image. Its form is: ； in, Indicates the first A learnable deep denoising prior operator for each stage enables the mapping from low-quality sequential stream images to high-fidelity sequential stream image representations.

4. The image compression and reconstruction method based on block modulation sequence stream compression according to claim 1, characterized in that, In step (5), the cross-attention enhancement module is implemented as follows: The deep features output from the previous stage are first concatenated with the initial features of the current stage along the channel dimension to obtain the cross-stage concatenated input features. ; Input features after cross-stage concatenation Perform global average pooling to obtain channel-level statistical descriptions, and generate attention weights through a layer-by-layer mapping function; Attention weight generation and enhanced output features Represented in the following form: ； in, Indicates global average pooling; and Represents a learnable linear mapping; It is a non-linear activation function; For the Sigmoid function; This is a channel-by-channel multiplication operation.

5. The image compression and reconstruction method based on block modulation sequence stream compression according to claim 1, characterized in that, In step (5), the Mamba module with multi-mode expansion is implemented as follows: For the input features of the Mamba module with multi-modal expansion, they are expanded in three modes: spatial horizontal expansion, spatial vertical expansion, and channel-dimensional expansion, resulting in sequences. The recursive calculation process within each mode can be described as follows: ； After discretization, its evolutionary update form is as follows: ； The corresponding features are output from the hidden state: ； in, It is a state vector; These are learnable parameters; To address the discrepancy in contribution among different modes, a set of learnable weighted parameters is further introduced. This is used to dynamically adjust the fusion priority of different modes. The fused output is represented as: ； in, express The three unfolding modes of the data.

6. The image compression and reconstruction method based on block modulation sequence stream compression according to claim 1, characterized in that, In step (5), the channel attention module is implemented as follows: The input features are first rearranged according to the channel direction, flattening the two-dimensional space into a one-dimensional sequence representation, and then the query is obtained through linear mapping. Q ,key K AND value V Three tensors; the input feature tensor is represented as Then its linear transformation is expressed as: ； ； ； in, These are the parameters for network learning.

7. Then, the corresponding channel correlation matrix is ​​weighted to obtain the attention map, and based on this, the feature selection and fusion are completed, resulting in enhanced channel features. Specifically, it is expressed as: ； ； in, To normalize the scale and stabilize the attention weight distribution, a residual fusion method is finally used to integrate the enhanced channel features. Compared with the original input features Element-wise superposition is performed to form a stable enhanced output. To achieve non-local expression enhancement at the semantic level: 。 8. An image compression and reconstruction system based on block modulation sequence stream compression, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the image compression and reconstruction method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the image compression and reconstruction method according to any one of claims 1-6.

Images