Image compressed sensing reconstruction method based on double-domain optimization and multi-scale feature fusion

Abstract

The invention discloses an image compressed sensing reconstruction method based on double-domain optimization and multi-scale feature fusion, and the method comprises the following steps: (1) constructing a training data set based on a double-domain optimization and multi-scale feature fusion network, and generating final fusion features through an adaptive weighting mechanism; (2) constructing a compressed sensing reconstruction model of multi-scale feature fusion based on double-domain optimization; (3) constructing an image sampling sub-network and an initial reconstruction sub-network; (4) constructing a deep double-domain optimization and multi-scale reconstruction network; (5) reconstructing the image by the compressed sensing reconstruction model according to the obtained image information, and outputting a reconstructed image; and (6) evaluating the compressed sensing reconstruction model by using the existing image data, carrying out subjective quality comparison on the reconstructed image and the original image, constructing a loss function of image reconstruction, and carrying out network training on the compressed sensing reconstruction model to further optimize the compressed sensing reconstruction model. The method provided by the invention can realize efficient image reconstruction effect and precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of image quality enhancement, and specifically provides an image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion. Background Art

[0002] Compressed sensing (CS) is a technology designed to find sparse solutions in underdetermined linear systems. Its core is to perform compression operations while sampling data. CS overcomes the limitations of the Nyquist sampling theorem in traditional signal processing, enabling efficient data acquisition and reconstruction at extremely low sampling rates. It has attracted considerable attention in the field of image processing. Specifically, CS uses sparse representation and random sampling techniques to reconstruct high-quality results close to the original image from measurements at sampling rates far below the Nyquist rate. This capability significantly reduces data storage and transmission costs, providing a new solution for resource-constrained scenarios.

[0003] Currently, CS technology has been widely used in many fields, including image compression, single-pixel cameras, medical imaging, snapshot compression imaging, remote sensing, radar image acquisition, and wireless remote monitoring. Its advantage is that it can restore high-quality signals with very few measurements, providing a new path for efficient data processing.

[0004] However, existing image compression sensing methods often utilize a single domain (e.g., pixel or feature domain) for processing, neglecting the synergistic nature of images between the pixel and feature domains. This limitation not only restricts image reconstruction accuracy but also makes it difficult to recover image details at low sampling rates. Furthermore, traditional methods fail to fully exploit the multi-scale feature information of images, further limiting reconstruction performance. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an image compressed sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion. This method can achieve efficient image reconstruction effect and accuracy by deeply fusing the pixel domain and feature domain information of the image and introducing the attention mechanism into the multi-scale feature reconstruction stage.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for image compression sensing reconstruction based on dual-domain optimization and multi-scale feature fusion, comprising the following steps:

[0007] (1) Construct a training dataset based on dual-domain optimization and multi-scale feature fusion network, and input the original image data into the trained image compression sensing model. The image compression sensing model includes a dual-domain sampling module DDSM, a dual-domain fusion module DDFM and an attention-based multi-scale feature fusion module MFFM. The dual-domain sampling module DDSM samples the original image in blocks through a random measurement matrix. The sampling content includes pixel domain and feature domain, and generates sparse sub-sampling features. The sub-sampling features are processed by the dual-domain fusion module DDFM, where the high-resolution branch extracts local detail features through residual convolution, and the low-resolution branch extracts global features through convolution and downsampling. The multi-scale features of different branches are fused in DDFM respectively. The fused features are subjected to channel attention and spatial attention by the multi-scale feature fusion module MFFM, and the final fusion features are generated through an adaptive weighting mechanism.

[0008] (2) Construct a compressed sensing reconstruction model based on multi-scale feature fusion based on dual-domain optimization;

[0009] (3) Constructing an image sampling subnetwork and an initial reconstruction subnetwork, continuing to sample images and initially optimizing the compressed sensing reconstruction model in step (2);

[0010] (4) Constructing a deep dual-domain optimization and multi-scale reconstruction network, the compressed sensing reconstruction model in step (2) optimizes and enhances the image information obtained by the dual-domain sampling module DDSM through pixel domain and feature domain sampling, and the multi-scale feature fusion module MFFM denoises the image information of different resolutions;

[0011] (5) The compressed sensing reconstruction model reconstructs the image based on the obtained image information and outputs the reconstructed image;

[0012] (6) Use existing image data to evaluate the compressed sensing reconstruction model, compare the subjective quality of the reconstructed image with the original image, construct the loss function of image reconstruction, and perform network training on the compressed sensing reconstruction model to further optimize the compressed sensing reconstruction model.

[0013] The sampling subnet achieves linear sampling of the image through the DDSM. The DDSM module combines information from the pixel domain and feature domain and utilizes a convolutional neural network to perform multi-dimensional sampling of the image, significantly improving the diversity of measurement content and information density. Unlike traditional methods that rely solely on pixel-domain sampling, the dual-domain sampling module provides more fine-grained measurement information. The initial reconstruction subnet generates a preliminary reconstructed image in the measurement domain, and the deep reconstruction subnet uses the DDFM module to further fuse pixel-domain and feature-domain information to optimize the image reconstruction process. The MFFM module introduces an attention mechanism to capture the correlation between features of different resolutions, effectively improving the model's reconstruction capabilities while also achieving denoising. The synergistic effect of the two modules significantly improves reconstruction quality and ensures the accurate restoration of image details.

[0014] The proximal gradient descent algorithm is used to solve the pixel domain and feature domain joint sampling model constructed in step (2) and the multi-scale compressed sensing reconstruction model with dual-domain fusion optimization.

[0015] Preferably, when the dual-domain sampling module DDSM subsamples the original image using a random measurement matrix, the measurement matrix is generated by Xavier initialization, and an initial reconstructed feature map is generated in combination with an initialization back-projection operation during the sampling process.

[0016] Preferably, when the dual-domain sampling module DDSM performs block sampling, the original image is decomposed into multiple sub-blocks of fixed size. Each sub-block is independently compressed and measured to form local sampling features. The sub-blocks achieve consistency by sharing the measurement matrix, and the local features are optimized in the subsequent fusion stage.

[0017] Preferably, the dual-domain fusion module DDFM performs feature fusion on the features of the high-resolution and low-resolution branches through channel splicing, and the fused features are further compressed into channel dimensions through a 1×1 convolutional layer to reduce redundant information.

[0018] Preferably, the convolution operation of the low-resolution branch realizes deep feature extraction through multi-layer stacking, and the downsampling process uses a convolution operation with a stride of 2 to ensure the integrity of the global information in the low-resolution features.

[0019] Preferably, the multi-scale feature fusion module MFFM includes:

[0020] The channel attention module and the spatial attention module are used to improve the feature importance of different channels and the feature expression ability of spatial positions respectively;

[0021] Adaptive weighting mechanism dynamically adjusts the weights of different feature fusion methods to select the optimal feature fusion strategy in a specific scenario;

[0022] The attention-based multi-scale feature fusion module (MFFM) performs hierarchical fusion of the input multi-scale features, refines the fusion results layer by layer, and improves the feature representation capability in a step-by-step optimization manner.

[0023] Preferably, the channel attention module uses global average pooling to generate global statistical information and adjusts the weights of each channel through a multi-layer perceptron; the spatial attention module combines maximum pooling and average pooling operations to generate a spatial attention map, and multiplies it element-by-element with the input features to achieve spatial enhancement.

[0024] Preferably, it also includes an image acquisition module for collecting the original image to be compressed sensing reconstructed; an image compression module for compressing and sampling the original image through a dual-domain sampling module DDSM to generate sparse sampling features; and an image reconstruction module for processing the sampling features through a dual-domain fusion module DDFM and an attention-based multi-scale feature fusion module MFFM, extracting multi-scale fusion features and outputting a reconstructed image.

[0025] Preferably, the method further comprises a dynamic measurement matrix generation module, which dynamically adjusts the initialization parameters of the measurement matrix according to the resolution and compression ratio requirements of the input image.

[0026] The training dataset of the compressed sensing reconstruction model in step (1) uses 400 images from the training and test sets of the BSDS500 dataset, and uses data augmentation technology to crop the above images into 89,600 96*96 pixel sub-images. The model trains five networks on a set of CS ratios of {10%, 25%, 30%, 40%, 50%}. During the training process, a batch size of 32 image blocks is used for training, and the Adam optimizer is used to optimize the network parameters. The initial learning rate is set to 2×10 -4 , use the cosine annealing strategy to update the learning rate and reduce it to 5×10 -5 The network was trained for a total of 120 epochs. Except for the input and output layers, the number of output channels in other layers was set to 16 by default. Three publicly available datasets were used as test datasets, including Set11, BSDS68, and Urban100. Color images were processed in YCbCr space and evaluated on the Y channel. Two commonly used image evaluation metrics, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), were used to evaluate the CS reconstruction results.

[0027] When constructing the image sampling subnetwork and the initial reconstruction subnetwork in step (3), first input the image X and decompose it into non-overlapping image blocks use Get the observation value of each image block Here we use large-step convolution operations and learnable sampling matrices to implement dual-domain sampling for compressed sensing. Get the initialized image block Here, a sampling convolution operation is used to initialize the image, and then the pixel shuffle operation is used to reconstruct the initial image to obtain the preliminary reconstructed image X. init , the convolution operation and pixel shuffling operation here can be simply referred to as the initial reconstruction subnetwork.

[0028] In step (6), the loss function of image reconstruction is constructed. The total target loss consists of two parts: the image reconstruction loss and the orthogonal constraint of the sampling matrix. The mean square error (MSE) is used as the loss function to train the network.

[0029] The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion provided by the present invention has the following beneficial effects:

[0030] 1. Through dual-domain sampling and fusion mechanisms, this method achieves higher information transfer efficiency during image reconstruction. In particular, it can fully utilize the diversity of sampled content at low sampling rates. Compared with traditional methods, this method not only excels in quantitative metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), but also restores clearer edges and details in images, significantly improving the visual quality. This makes this method more practical in image compressed sensing applications, especially in image compression, medical imaging, single-pixel cameras, and other fields.

[0031] 2. The introduced multi-scale feature fusion module, combined with the attention mechanism, can accurately capture local features and global information of images at different scales. By weighting features of different resolutions, the model network can better restore image details and textures, especially in complex image scenes, and effectively suppress blurring. This has significant advantages for tasks such as detail reconstruction and noise removal, and improves the network model's adaptability to various image structures.

[0032] 3. The introduction of an adaptive sampling matrix optimizes the sampling process by combining image prior information and structural features, enabling the network model to recover high-quality images with less measurement data. Compared with traditional methods, this method has improved computational complexity and provides faster reconstruction speed through a deep learning-based architecture and GPU acceleration. It is particularly suitable for resource-constrained application scenarios such as real-time image processing and embedded systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the model network architecture of the present invention;

[0034] Figure 2 Schematic diagram of the DDSM, DDFM and MFFM modules of the present invention;

[0035] Figure 3 This is a comparison chart of the subjective quality of the "Boats" image reconstructed on the Set11 dataset using the method of the present invention and the six closest compressed sensing reconstruction methods;

[0036] Figure 4 This is a comparison chart of the subjective quality of the image “image_019” reconstructed on the Urban100 dataset using the method of the present invention and the six closest compressed sensing reconstruction methods. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0038] See also Figures 1 to 4 The present invention provides a technical solution: an image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion, comprising the following steps:

[0039] (1) Construct a training dataset based on dual-domain optimization and multi-scale feature fusion network, and input the original image data into the trained image compression sensing model. The image compression sensing model includes a dual-domain sampling module DDSM, a dual-domain fusion module DDFM and an attention-based multi-scale feature fusion module MFFM. The dual-domain sampling module DDSM samples the original image in blocks through a random measurement matrix. The sampling content includes pixel domain and feature domain, and generates sparse sub-sampling features. The sub-sampling features are processed by the dual-domain fusion module DDFM, where the high-resolution branch extracts local detail features through residual convolution, and the low-resolution branch extracts global features through convolution and downsampling. The multi-scale features of different branches are fused in DDFM respectively. The fused features are subjected to channel attention and spatial attention by the multi-scale feature fusion module MFFM, and the final fusion features are generated through an adaptive weighting mechanism.

[0040] (2) Construct a compressed sensing reconstruction model based on multi-scale feature fusion based on dual-domain optimization;

[0041] (3) Constructing an image sampling subnetwork and an initial reconstruction subnetwork, continuing to sample images and initially optimizing the compressed sensing reconstruction model in step (2);

[0042] (4) Constructing a deep dual-domain optimization and multi-scale reconstruction network, the compressed sensing reconstruction model in step (2) optimizes and enhances the image information obtained by the dual-domain sampling module DDSM through pixel domain and feature domain sampling, and the multi-scale feature fusion module MFFM denoises the image information of different resolutions;

[0043] (5) The compressed sensing reconstruction model reconstructs the image based on the obtained image information and outputs the reconstructed image;

[0044] (6) Use existing image data to evaluate the compressed sensing reconstruction model, compare the subjective quality of the reconstructed image with the original image, construct the loss function of image reconstruction, and perform network training on the compressed sensing reconstruction model to further optimize the compressed sensing reconstruction model.

[0045] The implementation steps of the sampling network, initial reconstruction network and deep multi-scale feature reconstruction network of the present invention are as follows:

[0046] Step (1) Construction of training dataset and test dataset

[0047] Before training the model network, the BSDS500 dataset was used as the training dataset. 400 of the images were cropped into 89,600 sub-images using data augmentation techniques (such as rotation, flipping, and cropping). All sub-images were sized 96 × 96. The Set11, BSDS68, and Urban100 datasets were used as test datasets to evaluate the network's reconstruction performance. The test image sizes were consistent with the task requirements.

[0048] Step (2) Sampling network

[0049] The goal of the sampling network is to transform the input image into a sparse measurement vector, which includes the following processes:

[0050] Training generation of sampling matrices

[0051] Sampling Matrix It is obtained by training the network and initialized using Xavier initialization, which has good distribution characteristics and ensures the fidelity and sparsity of the signal during the measurement process;

[0052] Measurement operation

[0053] Use convolution operation to simulate the sampling process, input image Sampled matrix Get the measurement vector Y, Y = ΦX, sampling result is the sparsely represented measurement data, where M is the sampled dimension, determined by the sampling rate. In the implementation, the convolution kernel size and stride are adjusted through the block design of Φ, so that the measurement data can efficiently encode the key information of the input image;

[0054] Step (3) Initial reconstruction of the network

[0055] The initial reconstruction network converts the obtained measurement vector Y into a preliminary reconstructed image. The steps are as follows: transpose the sampling matrix Φ to obtain Φ T , for inverse transformation: X init =Φ T Y, through the convolution layer to X init Perform feature extraction and generate the initial image block. The design of this convolution layer is consistent with the structure of the sampling matrix and is used to restore the spatial details lost during the sampling process;

[0056] Use the pixel shuffle operation to reorganize the initial image blocks and restore the low-resolution features to the high-resolution initial reconstructed image: X shuffle =PixelShuffle(X init ), and finally obtain the preliminary reconstructed image X init , whose resolution is consistent with the input image and serves as the input of the subsequent deep multi-scale reconstruction network;

[0057] Step (4) Deep multi-scale reconstruction network

[0058] The deep multi-scale reconstruction network further optimizes the initial reconstructed image to generate high-quality reconstruction results. It mainly includes the following modules:

[0059] Multi-resolution feature extraction

[0060] Use parallel high-resolution and low-resolution branches:

[0061] The high-resolution branch extracts local detail features through the residual convolution module (ResBlock);

[0062] The low-resolution branch extracts global features through convolution downsampling with a stride of 2, and then upsamples to restore the resolution through bilinear interpolation. The features of the two branches are fused through channel splicing to generate multi-scale features.

[0063] Feature enhancement with attention mechanism

[0064] The Channel Attention module generates feature channel weights through global average pooling, highlighting important channels. The Spatial Attention module combines max pooling and average pooling to generate a spatial attention map, enhancing the representation of features at specific locations. This attention mechanism further refines the representation of multi-scale features and improves reconstruction performance.

[0065] Feature Fusion Module (MFFM)

[0066] The adaptive weighted feature fusion module performs weighted fusion of multi-scale features, combines the advantages of local and global features, and generates the final fusion feature: F final =w1·F local +w2·F global , where w1 and w2 are adaptive weights obtained through learning.

[0067] The output layer reconstructs the image

[0068] The fused features are input into the last layer of convolutional network to generate high-quality reconstructed images, and the fidelity of image details is improved through residual connections;

[0069] Step (5) Loss Function

[0070] The loss function consists of two parts: the image reconstruction loss using mean squared error (MSE) and the row orthogonality constraint loss of the sampling matrix.

[0071] Parameter configuration for network training

[0072] First, 400 images from the training and test sets of the BSDS500 dataset were used. These images were cropped into 89,600 96×96 pixel sub-images using data augmentation techniques. The proposed method was then implemented using the deep learning framework PyTorch. The model was trained for a total of 120 epochs, and the Adam optimizer was used to update the network parameters. The initial learning rate was set to 2×10 -4 , using the cosine annealing strategy to reduce it to 5×10 -5 , the warm-up period is 3; during the training process of the network designed by the present invention, the number of channels of the convolutional layer is set to 16, the batch size is set to 32, the number of stages is set to 16, and the size of the convolution kernel is set to 3×3; in order to better evaluate the performance of the proposed method, three widely used benchmark datasets are used for testing, including Set11, BSD68, and Urban100; finally, the trained model is evaluated using two objective indicators, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and the reconstructed images are also compared for subjective quality;

[0073] Network training

[0074] In the image compression sensing task, the goal of the model is to reconstruct the original image X from the measurement value Y and make the reconstructed image X -dr As close as possible to the real image X. To achieve this goal, the mean square error (MSE) is selected as the main loss function. MSE not only has good differentiability, but also has become a commonly used loss function in many compressed sensing methods based on deep learning due to its fast convergence speed. Specifically expressed as

[0075]

[0076] In order to improve the recovery performance of the sampling matrix and promote the generalization ability of the model, the row orthogonality constraint of the sampling matrix is introduced into the loss function. This constraint avoids overfitting and improves the reconstruction quality by controlling the orthogonality of the sampling matrix. The specific form is

[0077]

[0078] The final loss function combines the above two parts, where μ is a hyperparameter that controls the balance between the regularization term and the reconstruction error. The overall form of the loss function is

[0079]

[0080] Among them, L mse is the mean square error between the reconstructed image and the original image, L orth is the orthogonality constraint of the sampling matrix, is the X of the i-th image i Final output reconstruction, N b is the batch size, I is the identity matrix, and μ is set to 0.01 during network training.

[0081] Objective quality comparison of network test results:

[0082] To validate the effectiveness of our method, we compared our proposed method for compressed image sensing reconstruction based on dual-domain optimization and multi-scale feature fusion with seven state-of-the-art deep learning-based methods: ISTA-Net+, DPA-Net, AMP-Net, DPC-DUN, COAST, MADUN, and MDGF-Net+. Tables 1-3 show the average PSNR / SSIM comparisons of these methods on the Set11 dataset. Table 2 shows the average PSNR / SSIM comparisons of these methods on the Urban100 dataset, and Table 3 shows the average PSNR / SSIM comparisons of these methods on the BSDS68 dataset. As can be seen from Table 3, the objective metrics of the reconstructed images by our method on the BSDS68 dataset far outperform those of the other seven compressed image sensing reconstruction methods, demonstrating the advantages of our method. However, on the Set11 and Urban100 datasets, our method exhibits lower PSNR and SSIM than MADUN at certain sampling rates. This is because MADUN utilizes more training stages and parameters to achieve a slight performance improvement. Compared to MADUN, our method achieves a better balance between performance and network complexity. At the same time, the average PSNR and SSIM indicators of the proposed method are better than those of other methods.

[0083] Table 1:

[0084]

[0085]

[0086] Table 2:

[0087]

[0088] Table 3:

[0089]

[0090] Visual quality comparison of network test results:

[0091] In order to verify the effectiveness of the method of the present invention, the image compressed sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion proposed in the present invention is compared with ISTA-Net+, AMP-Net, DPC-DUN, MADUN, COAST and MDGF-Net+. Figure 3 and Figure 4 The following figure shows a visual comparison of images from the Set11 and Urban100 datasets. As can be seen from these two figures, when the sampling rates are 25% and 30%, the images restored by our method retain more detailed information and are clearer than those restored by other compressed sensing reconstruction methods, fully demonstrating the superiority of our method.

[0092] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion, characterized by: The following steps are involved: (1) Construct a training dataset based on dual-domain optimization and multi-scale feature fusion network, and input the original image data into the trained image compression sensing model. The image compression sensing model includes a dual-domain sampling module DDSM, a dual-domain fusion module DDFM and an attention-based multi-scale feature fusion module MFFM. The dual-domain sampling module DDSM samples the original image in blocks through a random measurement matrix. The sampling content includes pixel domain and feature domain, and generates sparse sub-sampling features. The sub-sampling features are processed by the dual-domain fusion module DDFM, where the high-resolution branch extracts local detail features through residual convolution, and the low-resolution branch extracts global features through convolution and downsampling. The multi-scale features of different branches are fused in DDFM respectively. The fused features are subjected to channel attention and spatial attention by the multi-scale feature fusion module MFFM, and the final fusion features are generated through an adaptive weighting mechanism. (2) Construct a compressed sensing reconstruction model based on multi-scale feature fusion based on dual-domain optimization; (3) Constructing an image sampling subnetwork and an initial reconstruction subnetwork, continuing to sample images and initially optimizing the compressed sensing reconstruction model in step (2); (4) Constructing a deep dual-domain optimization and multi-scale reconstruction network, the compressed sensing reconstruction model in step (2) optimizes and enhances the image information obtained by the dual-domain sampling module DDSM through pixel domain and feature domain sampling, and the multi-scale feature fusion module MFFM denoises the image information of different resolutions; (5) The compressed sensing reconstruction model reconstructs the image based on the obtained image information and outputs the reconstructed image; (6) Use existing image data to evaluate the compressed sensing reconstruction model, compare the subjective quality of the reconstructed image with the original image, construct the loss function of image reconstruction, and perform network training on the compressed sensing reconstruction model to further optimize the compressed sensing reconstruction model.

2. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 1, characterized in that: When the dual-domain sampling module DDSM subsamples the original image using a random measurement matrix, the measurement matrix is generated by Xavier initialization, and an initial reconstructed feature map is generated in combination with an initialization back-projection operation during the sampling process.

3. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 2, characterized in that: When the dual-domain sampling module (DDSM) performs block sampling, the original image is decomposed into multiple sub-blocks of fixed size. Each sub-block is independently compressed and measured to form local sampling features. The sub-blocks achieve consistency by sharing the measurement matrix, and the local features are optimized in the subsequent fusion stage.

4. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 3, characterized in that: The dual-domain fusion module (DDFM) fuses the features of the high-resolution and low-resolution branches through channel splicing. The fused features are further compressed through a 1×1 convolutional layer to reduce redundant information.

5. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 4, characterized in that: The convolution operation of the low-resolution branch achieves deep feature extraction through multi-layer stacking, and the downsampling process uses a convolution operation with a stride of 2 to ensure the integrity of the global information in the low-resolution features.

6. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 5, characterized in that: The multi-scale feature fusion module MFFM includes: The channel attention module and the spatial attention module are used to improve the feature importance of different channels and the feature expression ability of spatial positions respectively; Adaptive weighting mechanism dynamically adjusts the weights of different feature fusion methods to select the optimal feature fusion strategy in a specific scenario; The attention-based multi-scale feature fusion module (MFFM) performs hierarchical fusion of the input multi-scale features, refines the fusion results layer by layer, and improves the feature representation capability in a step-by-step optimization manner.

7. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 6, characterized in that: The channel attention module uses global average pooling to generate global statistical information and adjusts the weights of each channel through a multi-layer perceptron; the spatial attention module combines maximum pooling and average pooling operations to generate a spatial attention map, and multiplies it element-by-element with the input features to achieve spatial enhancement.

8. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 7, characterized in that: It also includes an image acquisition module for acquiring the original image to be compressed sensing reconstructed; an image compression module for compressing and sampling the original image through a dual-domain sampling module DDSM to generate sparse sampling features; The image reconstruction module is used to process the sampled features through the dual-domain fusion module DDFM and the attention-based multi-scale feature fusion module MFFM, extract multi-scale fusion features and output the reconstructed image.

9. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 8, characterized in that: The invention also includes a dynamic measurement matrix generation module, which dynamically adjusts the initialization parameters of the measurement matrix according to the resolution and compression ratio requirements of the input image.

10. The image compression sensing reconstruction method based on dual-domain optimization and multi-scale feature fusion according to claim 9, characterized in that: The loss function for image reconstruction is constructed by taking the mean square error L between the reconstructed image and the original image mse and the orthogonality constraint L of the sampling matrix orth , then the final loss function L combines L mse and L mse , and introduce the hyperparameter μ to control the balance between the regularization term and the reconstruction error. The specific formula is as follows: Among them, L mse is the mean square error between the reconstructed image and the original image, L orth is the orthogonality constraint of the sampling matrix, is the X of the i-th image i Final output reconstruction, N b is the batch size, I is the identity matrix, and μ is set to 0.01 during network training.

Images