Image compressed sensing reconstruction methods, systems, equipment and media

Abstract

This invention discloses an image compressed sensing reconstruction method, system, device, and medium. The method includes: using a preset first denoiser model to degrade the original image to obtain a degraded image; using an improved deep learning-based approximate message passing algorithm to compress and reconstruct the degraded image to obtain the image compressed sensing reconstruction result; wherein, the improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; both the preset first denoiser model and the preset second denoiser model are grayscale image Gaussian noise denoiser network models based on image prior modeling; this invention further refines the noise interval used by the algorithm by utilizing the noise statistical distribution and denoising capability curve of the denoiser, effectively improving the image reconstruction capability.

Description

Technical Field

[0001] This invention belongs to the field of digital image processing technology, and specifically relates to an image compressed sensing reconstruction method, system, device and medium. Background Technology

[0002] With the continuous development of information technology, the total amount of information is also increasing. The huge amount of information and the higher real-time requirements have put forward higher requirements for signal transmission and storage. As one of the main carriers of information, image and video interaction technology has greatly influenced the development of communication technology. Therefore, efficient image and video compression algorithms have gradually become a hot topic in academic and industrial research.

[0003] Compressed sensing theory is a novel image compression theory that states that when a signal can be sparsely represented, it is possible to sample at a frequency much lower than that required by the Nyquist sampling theorem and accurately recover the original signal. Due to the low sampling rate of compressed sensing images, it has wide applications in spaceborne satellite surveillance, CT image reconstruction, and military fields. Existing image reconstruction methods based on compressed sensing theory can be divided into traditional compressed sensing reconstruction algorithms and compressed sensing reconstruction algorithms based on deep neural networks.

[0004] Traditional compressed sensing reconstruction algorithms include AMP, NLRCS, and CoSaMP. These algorithms often require enormous computing power to predict and reconstruct images from underdetermined problems. Therefore, traditional compressed sensing reconstruction algorithms often suffer from high computational complexity and unstable reconstruction performance.

[0005] Compressed sensing reconstruction algorithms based on deep neural networks, including LDAMP, ADMM-Net, and OPINE-Net, utilize neural networks to replace some complex prediction parts in traditional compressed sensing reconstruction algorithms, achieving a significant improvement in compressed sensing reconstruction performance. However, at extremely low sampling rates, compressed sensing-based coded reconstruction algorithms suffer from low rate-distortion performance and poor reconstruction results compared to traditional image compression algorithms. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides an image compressed sensing reconstruction method, system, device, and medium to solve the technical problem that compressed sensing-based coded reconstruction algorithms have low rate-distortion performance and poor reconstruction effect compared to traditional image compression algorithms at extremely low sampling rates.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] This invention provides an image compressed sensing reconstruction method, comprising:

[0009] Using a preset first denoising model, the original image to be processed is degraded to obtain a degraded image;

[0010] An improved deep learning-based approximate message passing algorithm is used to compress and reconstruct the degraded image, resulting in the image compressed sensing reconstruction.

[0011] The improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; the preset first denoiser model and the preset second denoiser model are both grayscale image Gaussian noise denoiser network models based on image prior modeling.

[0012] Furthermore, the grayscale image Gaussian noise denoising network model based on image prior modeling includes several levels of noise models; wherein each level of noise model includes a first convolutional layer, a first multi-scale fusion module, a first supervised attention module, a second multi-scale fusion module, a channel merging module, a second supervised attention module, a second convolutional layer, an attention module, a first residual connection module, a third convolutional layer, and a second residual connection module; wherein the previous level noise model is a preprocessing model for the next level noise model, used to fine-tune the training of the next level noise model;

[0013] The first convolutional layer is used to perform convolution operations on the noisy grayscale image to obtain the initial feature map F. hea ;

[0014] The first multi-scale fusion module is used to process the initial feature map F. head Feature extraction is performed to obtain feature map F. m1 ;

[0015] The first supervised attention module is used to process the feature map F. m1 Feature calibration is performed, and the number of channels is changed to obtain the first-stage denoised image.

[0016] The second multi-scale fusion module is used to process the feature map F m1 Feature extraction is performed to obtain feature map F. m2 ;

[0017] The channel merging module is used to process the feature map F m1 and the feature map F m2 Perform channel merging to obtain the merged feature map F. m12 ;

[0018] The first supervised attention module is used to process the feature map F after channel merging. m12Feature calibration is performed, and the number of channels is changed to obtain the second-stage denoised image.

[0019] The second convolutional layer is used to process the feature map F after channel merging. m12 Feature extraction is performed to obtain feature map F. c2 ;

[0020] Attention module, used for the feature map F c2 Channel and spatial attention features are extracted to obtain feature map F. csa ;

[0021] The first residual connection module is used to process the initial feature map F. head and the feature map F csa The feature map F is obtained by summing the results. res1 ;

[0022] The third convolutional layer is used to process the feature map F. res1 Feature extraction is performed to obtain feature map F. c3 ;

[0023] The second residual connection module is used to connect the noisy grayscale image with the feature map F. c3 Summation is performed to obtain the third-stage denoised image.

[0024] The present invention also provides an image compressed sensing reconstruction system, including

[0025] The degradation module is used to perform degradation processing on the original image to be processed using a preset first denoising model to obtain a degraded image;

[0026] The reconstruction module is used to perform compressed reconstruction processing on the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compressed sensing reconstruction result.

[0027] The improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; the preset first denoiser model and the preset second denoiser model are both grayscale image Gaussian noise denoiser network models based on image prior modeling.

[0028] The present invention also provides an image compressed sensing reconstruction device, comprising:

[0029] Memory, used to store computer programs;

[0030] A processor is used to implement the steps of the image compressed sensing reconstruction method when executing the computer program.

[0031] The present invention also provides a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the image compressed sensing reconstruction method.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] This invention provides an image compressed sensing reconstruction method and system. It utilizes a grayscale image Gaussian noise denoiser network model based on image prior modeling to perform degradation processing on the original image to be processed. This degradation preprocessing reduces the interference of high-frequency details on low-frequency information during compressed sensing reconstruction, significantly improving the image reconstruction effect of the algorithm. Simultaneously, it improves the deep learning-based approximate message passing algorithm using the grayscale image Gaussian noise denoiser network model based on image prior modeling, and uses the improved algorithm for image compressed reconstruction. The noise intervals used by the algorithm are further refined using the noise statistical distribution and denoising capability curve of the denoiser, effectively improving the image reconstruction capability. The deep neural network architecture has wide applicability and can be used for various low-level computer vision tasks, such as image denoising, image dehazing, and image enhancement, showing broad application prospects.

[0034] Furthermore, the grayscale image Gaussian noise denoising network model based on image prior modeling, using the MWCNN network as a foundation, acquires and fuses local and local features of the image by designing different sub-networks. In the sub-network for acquiring local image features, a self-designed deformable attention module is added to each layer of feature extraction, thus achieving comprehensive utilization of various image features. Channel attention mechanism and spatial attention mechanism are added to the network, making the network pay more attention to useful spatial pixel information and channel information in the feature map and suppressing useless information in the feature map. Attached Figure Description

[0035] Figure 1 This is a structural diagram of the noise model in this invention;

[0036] Figure 2 This is a structural diagram of the multi-scale feature fusion module in this invention;

[0037] Figure 3 This is a structural diagram of the deformable attention module in this invention;

[0038] Figure 4 This is a structural diagram of the self-attention module in this invention;

[0039] Figure 5 This is a structural diagram of the supervised attention module in this invention;

[0040] Figure 6 This is a flowchart of the improved deep learning-based approximate message passing algorithm in the embodiment;

[0041] Figure 7 The image is the test image used in the embodiment; wherein, Figure 7 a is the test image Parrot. Figure 7 b represents the test image, House;

[0042] Figure 8 This is a comparative schematic diagram of the compressed sensing reconstruction test results of the test image Parrot in the embodiment;

[0043] Figure 9 This is a comparative schematic diagram of the compressed sensing reconstruction test results of the test image "House" in the embodiment. Detailed Implementation

[0044] To make the technical problems solved by the present invention, the technical solutions, and the beneficial effects clearer, the following specific embodiments provide a further detailed description of the present invention. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of the invention.

[0045] This invention provides an image compressed sensing reconstruction method, comprising the following steps:

[0046] Step 1: Using the preset first denoising model, the original image to be processed is degraded to obtain the degraded image.

[0047] Step 2: Using an improved deep learning-based approximate message passing algorithm, the degraded image is compressed and reconstructed to obtain the image compressed sensing reconstruction result; specifically, when performing the t-th layer iterative reconstruction calculation, the divergence of the t-th level noise model... for:

[0048]

[0049] The deviation z between the calculated reconstruction result of the t-th layer iterative reconstruction calculation and the actual measurement value. t for:

[0050]

[0051] The estimated standard deviation σ between the calculated reconstruction result and the actual measured value in the t-th layer iterative reconstruction calculation. t for:

[0052]

[0053]

[0054]

[0055] The iterative reconstruction calculation at layer t produces data b of standard normal distribution with size [1, N]:

[0056] b = rand(1, N)

[0057] Where x is the original image to be processed, x D Let y be the degraded image and x be the degraded image. D The measured values ​​after the measurement matrix; M is the number of measured values; N is the total number of pixels in the image, x t Let |x|| be the reconstructed data vector of layer t; ∞ Let b represent the infinite norm of x. H Let b be the transpose of b, and A be the measurement matrix. H To measure the transpose of the matrix, This represents a noise model.

[0058] In this invention, the improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; both the preset first denoiser model and the preset second denoiser model are grayscale image Gaussian noise denoising network models based on image prior modeling.

[0059] The grayscale image Gaussian noise denoising network model based on image prior modeling includes several levels of noise models; wherein, the previous level noise model is a preprocessing model for the next level noise model, used for fine-tuning the training of the next level noise model; as shown in the appendix. Figure 1 As shown, each level of the noise model includes a first convolutional layer, a first multi-scale fusion module, a first supervised attention module, a second multi-scale fusion module, a channel merging module, a second supervised attention module, a second convolutional layer, an attention module, a first residual connection module, a third convolutional layer, and a second residual connection module.

[0060] In this invention, the first convolutional layer is used to perform a convolution operation on a noisy grayscale image to obtain an initial feature map F. head The first multi-scale fusion module is used to process the initial feature map F. head Feature extraction is performed to obtain feature map F. m1 The first supervised attention module is used to process the feature map F. m1 Feature calibration is performed, and the number of channels is changed to obtain the first-stage denoised image. The second multi-scale fusion module is used to process the feature map F m1 Feature extraction is performed to obtain feature map F.m2 ; Channel merging module, used for merging the feature map F m1 and the feature map F m2 Perform channel merging to obtain the merged feature map F. m12 The first supervised attention module is used to process the feature map F after channel merging. m1 Feature calibration is performed, and the number of channels is changed to obtain the second-stage denoised image. The second convolutional layer is used to process the feature map F after channel merging. m12 Feature extraction is performed to obtain feature map F. c2 Attention module, used for processing the feature map F c2 Channel and spatial attention features are extracted to obtain feature map F. csa The first residual connection module is used to process the initial feature map F. head and the feature map F csa The feature map F is obtained by summing the results. res1 The third convolutional layer is used to process the feature map F. res1 Feature extraction is performed to obtain feature map F. c3 The second residual connection module is used to connect the noisy grayscale image with the feature map F. c3 Summation is performed to obtain the third-stage denoised image.

[0061] In this invention, the training process for each level of the noise model is as follows:

[0062] A training dataset is constructed, comprising several pairs of training data. Each pair of training data includes real data and noise data. The real data consists of clean grayscale images selected from the DIV2K and SIDD datasets, and the noise data consists of Gaussian noise of a preset noise level added to the clean grayscale images. Preferably, the level of each noise model is adapted to the noise level of the original image to be processed. The noise level of the original image to be processed is obtained by dividing the noise range of the original image to be processed using the standard deviation of Gaussian noise as the grading standard.

[0063] Normalize each training data point to obtain normalized training data.

[0064] The normalized training data is used as input to the current noise model, and the current noise model is trained based on the target loss function; wherein, the target loss function is based on the clean grayscale image and the first-stage denoised image. Second stage denoised image and the third-stage denoised image Calculated;

[0065] The target loss function is:

[0066]

[0067] Where MSE(*) is the mean squared error; x is a clean grayscale image; This is the first stage of denoising the image; For the second stage of denoising; This is the image for the third stage of denoising.

[0068] Once the current noise model is trained, it is used as a preprocessing model for the next level noise model, and the next level noise model is fine-tuned. This process is repeated to train all noise models.

[0069] In this invention, both the first multi-scale fusion module and its associated second multi-scale fusion module are multi-scale feature fusion modules; as shown in the appendix. Figure 2 As shown, the multi-scale feature fusion module includes a Transformer branch, a CNN branch, a first channel merging-convolution unit, a second channel merging-convolution unit, a third channel merging-convolution unit, a fourth channel merging-convolution unit, and a fifth channel merging-convolution unit.

[0070] The Transformer branch includes, in sequence, a first convolutional layer unit, a first self-attention module, a first discrete wavelet transform unit, a second convolutional layer unit, a second self-attention module, a second discrete wavelet transform unit, a third convolutional layer unit, a third self-attention module, a fourth convolutional layer unit, an inverse transform unit of the first discrete wavelet transform, a fourth self-attention module, a fifth convolutional layer unit, an inverse transform unit of the second discrete wavelet transform, a fifth self-attention module, a sixth convolutional layer unit, and a first output unit.

[0071] The first convolutional layer unit is used to process the initial feature map F. head Perform a convolution operation to obtain the feature map F. Trans-c7 The first self-attention module is used to process the feature map F. Trans-c7 Self-attention feature extraction is performed to obtain feature map F. Trans-sa1 The first discrete wavelet transform unit is used to process the feature map F. Trans-sa1 Discrete wavelet transform downsampling is performed to obtain the feature map F. Trans-dwt3 The second convolutional layer unit is used to process the feature map F. Trans-dwt3 Perform a convolution operation to obtain the feature map F. Trans-c8The second self-attention module is used to process the feature map F. Trans-c8 Self-attention feature extraction is performed to obtain feature map F. Trans-sa2 The second discrete wavelet transform unit is used to process the feature map F. Trans-sa Discrete wavelet transform downsampling is performed to obtain the feature map F. Trans-dwt4 The third convolutional layer unit is used to process the feature map F. Trans-dwt Perform a convolution operation to obtain the feature map F. Trans-c9 The third self-attention module is used to process the feature map F. Trans-c9 Self-attention feature extraction is performed to obtain feature map F. Trans-sa3 The fourth convolutional layer unit is used to process the feature map F. Trans-sa3 Perform a convolution operation to obtain the feature map F. Trans-c10 The inverse transform unit of the first discrete wavelet transform is used to transform the feature map F. Trans-c10 The feature map F is obtained by performing inverse discrete wavelet transform upsampling. Trans-iwt3 The fourth self-attention module is used to process the feature map F. Trans-iwt Self-attention feature extraction is performed to obtain feature map F. Trans-sa4 The fifth convolutional layer unit is used to process the feature map F. Trans-sa4 Perform a convolution operation to obtain the feature map F. Trans-c1 The inverse transform unit of the second discrete wavelet transform is used to transform the feature map F. Trans-c11 The feature map F is obtained by performing inverse discrete wavelet transform upsampling. Trans-twt4 The fifth self-attention module is used to process the feature map F. Trans-iwt Self-attention feature extraction is performed to obtain feature map F. Trans-sa The sixth convolutional layer unit is used to process the feature map F. Trans-sa5 Perform a convolution operation to obtain the output feature map F of the Transformer branch. Y2 The first output unit is used to process the output feature map F of the Transformer branch. Y2 Output the results.

[0072] In this invention, the CNN branch includes, in sequence, a seventh convolutional layer unit, a first deformable attention module, a third discrete wavelet transform unit, an eighth convolutional layer unit, a second deformable attention module, a fourth discrete wavelet transform unit, a ninth convolutional layer unit, a third deformable attention module, a tenth convolutional layer unit, an inverse transform unit of the third discrete wavelet transform, a fourth deformable attention module, an eleventh convolutional layer unit, an inverse transform unit of the fourth discrete wavelet transform, a fifth deformable attention module, a twelfth convolutional layer unit, and a second output unit.

[0073] The seventh convolutional layer unit is used to process the initial feature map F. head Perform a convolution operation to obtain the feature map F. CNN -; The first deformable attention module is used to process the feature map F CNN-c1 Edge features and attention features are extracted to obtain feature map F. CNN-dcs1 The third discrete wavelet transform unit is used to process the feature map F. CNN-dcs1 Discrete wavelet transform downsampling is performed to obtain the feature map F. CNN-dwt1 The eighth convolutional layer unit is used to process the feature map F. CNN-dwt1 Perform a convolution operation to obtain the feature map F. CNN- The second deformable attention module is used to process the feature map F. CNN-c2 Edge features and attention features are extracted to obtain feature map F. CNN-dcs2 The fourth discrete wavelet transform unit is used for the feature map F. CNN-dcs2 Discrete wavelet transform downsampling is performed to obtain the feature map F. CNN-dwt2 The ninth convolutional layer unit is used to process the feature map F. CNN-dwt2 Perform a convolution operation to obtain the feature map F. CNN-c3 The third deformable attention module is used to process the feature map F. CNN-c3 Edge features and attention features are extracted to obtain feature map F. CNN-dcs3 The tenth convolutional layer unit is used to process the feature map F. CNN-dcs3 Perform a convolution operation to obtain the feature map F. CNN-c4 The inverse transform unit of the third discrete wavelet transform is used to transform the feature map F. CNN-c The feature map F is obtained by performing inverse discrete wavelet transform upsampling. CNN-iwt1 The fourth deformable attention module is used to process the feature map F. CNN-iwt1 Edge features and attention features are extracted to obtain feature map F. CNN-dcs4 The eleventh convolutional layer unit is used to process the feature map F. CNN-dcs4 Perform a convolution operation to obtain the feature map F. CNN-c5 The inverse transform unit of the fourth discrete wavelet transform is used to process the feature map F. CNN-c5 The feature map F is obtained by performing inverse discrete wavelet transform upsampling. CNN-iwt The fifth deformable attention module is used to process the feature map F. CNN-iwt2 Edge features and attention features are extracted to obtain feature map F. CNN-dcs5 The twelfth convolutional layer unit is used to process the feature map F. CNN-dcs5Perform convolution operations to obtain the output feature map F of the CNN branch. Y1 The second output unit is used to process the output feature map F of the CNN branch. Y1 Output the results.

[0074] The outputs of the first self-attention module and the first deformable attention module are both connected to the input of the first channel merging-convolution unit, and the output of the first channel merging-convolution unit is connected to the input of the third discrete wavelet transform unit. The first channel merging-convolution unit is used to perform channel merging operation on the output of the first self-attention module and the output of the first deformable attention module, and to perform convolution operation on the merged result to halve the number of channels in the merged result, and use it as the input of the third discrete wavelet transform unit.

[0075] The outputs of the second self-attention module and the second deformable attention module are both connected to the input of the second channel merging-convolution unit, and the output of the second channel merging-convolution unit is connected to the input of the fourth discrete wavelet transform unit. The second channel merging-convolution unit is used to perform channel merging operation on the output of the second self-attention module and the output of the second deformable attention module, and to perform convolution operation on the merged result to halve the number of channels in the merged result, and use it as the input of the fourth discrete wavelet transform unit.

[0076] The outputs of the third self-attention module and the third deformable attention module are both connected to the input of the third channel merging-convolution unit, and the output of the third channel merging-convolution unit is connected to the input of the tenth convolutional layer unit. The third channel merging-convolution unit is used to perform channel merging operations on the outputs of the third self-attention module and the third deformable attention module, and to perform convolution operations on the merged results to halve the number of channels in the merged results, which are then used as the input of the tenth convolutional layer unit.

[0077] The outputs of the fourth self-attention module and the fourth deformable attention module are both connected to the input of the fourth channel merging-convolution unit, and the output of the fourth channel merging-convolution unit is connected to the input of the eleventh convolutional layer unit. The fourth channel merging-convolution unit is used to perform channel merging operations on the outputs of the fourth self-attention module and the fourth deformable attention module, and to perform convolution operations on the merged result to halve the number of channels in the merged result, which is then used as the input of the eleventh convolutional layer unit.

[0078] The outputs of the fifth self-attention module and the fifth deformable attention module are both connected to the input of the fifth channel merging-convolution unit, and the output of the fifth channel merging-convolution unit is connected to the input of the twelfth convolutional layer unit. The fifth channel merging-convolution unit is used to perform channel merging operations on the outputs of the fifth self-attention module and the fifth deformable attention module, and to perform convolution operations on the merged result to halve the number of channels in the merged result, which is then used as the input of the twelfth convolutional layer unit.

[0079] In this invention, the first deformable attention module, the second deformable attention module, the third deformable attention module, the fourth deformable attention module, and the fifth deformable attention module are all deformable attention modules; as shown in the appendix. Figure 3 As shown, the deformable attention module includes a deformable convolutional unit, a thirteenth convolutional layer unit, a spatial attention unit, a channel attention unit, a channel merging unit, a fourteenth convolutional layer unit, and a third residual connection module.

[0080] The deformable convolutional unit is used to process the feature map F CNN-c Edge feature extraction is performed to obtain feature map F. CNN-dc The thirteenth convolutional layer unit is used to process the feature map F. CNN Perform a convolution operation to obtain the feature map F. CNN The spatial attention unit is used to process the feature map F. CNN-c Perform pixel domain weight redistribution to obtain feature map F CNN The channel attention unit is used to process the feature map F. CNN-c13 Perform a feature channel weight reallocation operation to obtain the feature map F. CNN-ca The channel merging unit is used to process the feature map F. CNN-s and the feature map F CNN-ca Perform inter-channel merging operations to obtain feature map F. CNN-c The fourteenth convolutional layer unit is used to process the feature map F. CNN-csa Perform a convolution operation to obtain the feature map F. CNN-c14 The third residual connection module is used to process the feature map F. CNN-c14 and the feature map F CNN-c13 Perform residual connection operations to obtain the output features F of deformable attention. CNN-dcsa .

[0081] In this invention, the first self-attention module, the second self-attention module, the third self-attention module, the fourth self-attention module, and the fifth self-attention module all employ self-attention modules; as shown in the appendix. Figure 4 As shown, the self-attention module includes a fifteenth convolutional layer unit, a first reconstruction unit, a second reconstruction unit, a third reconstruction unit, a first matrix inner product unit, a first normalization unit, a second matrix inner product unit, and a fourth reconstruction unit.

[0082] The fifteenth convolutional layer unit is used to process the feature map F Trans-c7 Perform feature extraction to obtain feature map F Trans Feature map F Trans and feature map F Trans The first reconstruction unit is used to reconstruct the feature map F. Trans Perform a reconstruction and unfolding operation to expand the 3D feature map into a 2D matrix form, thus obtaining the feature map. The second reconstruction unit is used to reconstruct the feature map F. Trans Perform a reconstruction and unfolding operation to expand the 3D feature map into a 2D matrix form, thus obtaining the feature map. The third reconstruction unit is used to reconstruct the feature map F. Trans Perform a reconstruction and unfolding operation to expand the 3D feature map into a 2D matrix form, thus obtaining the feature map. The first matrix inner product unit is used to process the feature map. and the feature map Perform the matrix inner product operation to obtain the feature map. The first normalization unit is used to normalize the feature map. Normalization is performed to obtain the feature map F. Trans-Sig The second matrix inner product unit is used to process the feature map F. Trans-Sig and the feature map Perform the matrix inner product operation to obtain the feature map. The fourth reconstruction unit is used to reconstruct the feature map. A reconstruction and merging operation is performed to merge the two-dimensional feature maps into a three-dimensional feature matrix, resulting in feature map F. Trans-QKV ; wherein, the feature map F Trans-QKV This refers to the output features of the self-attention module.

[0083] In this invention, the first supervised attention module and the second supervised attention module have the same structure and the same operation process type; as shown in the appendix. Figure 5As shown, taking the first supervised attention module as an example, the supervised attention module includes a sixteenth convolutional layer unit, a seventeenth convolutional layer unit, a fourth residual connection module, an eighteenth convolutional layer unit, a second normalization unit, a matrix multiplication unit, and a fourth residual connection unit.

[0084] The sixteenth convolutional layer unit is used to process the feature map F m1 Perform feature extraction to obtain feature map F c16 The seventeenth convolutional layer unit is used to process the feature map F. m1 Perform a feature channel number transformation operation to obtain the noisy image F. noise The fourth residual connection module is used to process the noisy image F. noise The original noisy image is then subjected to a residual connection operation to obtain the first-stage denoised image. The eighteenth convolutional layer unit is used to process the denoised image. Perform a feature channel number transformation operation to obtain the feature map F. c18 The second normalization unit is used to normalize the feature map F. c18 Normalization is performed to obtain the attention feature map F. Sig The matrix dot product unit is used to multiply the feature map F. c18 With the feature map F c1 Perform a matrix dot product operation to obtain the feature map F. dot The fourth residual connection unit is used to connect the feature map F m1 With the feature map F dot Perform residual connection operations to obtain the output F of the supervised attention module. SAM .

[0085] In this invention, the convolution kernels of the first, second, and third convolutional layers are all 1×1; the convolution kernels of the first to sixteenth convolutional layer units are all 3×3, and the ReLU activation function is used after convolution.

[0086] The image compressed sensing reconstruction method described in this invention is based on the MWCNN network. It acquires and fuses local and local features of the image by designing different sub-networks. In the sub-network for acquiring local image features, a self-designed deformable attention module is added to each layer of the feature extraction process, achieving comprehensive utilization of various image features. Finally, channel attention and spatial attention mechanisms are added to the network, making the network focus more on useful spatial pixel information and channel information in the feature map and suppressing useless information in the feature map. The proposed deep neural network structure is applied to the depth-based denoising approximate message passing algorithm LDAMP, and the denoising... The noise statistical distribution and denoising capability curve of the device further refine the noise range used by the algorithm, thereby improving the image reconstruction capability of the algorithm. The deep neural network architecture proposed in this invention has wide applicability and can be used for a variety of low-level computer vision tasks, such as image denoising, image dehazing, and image enhancement, and has broad application prospects. The proposed gray-level Gaussian denoising network structure with refined noise distribution is applied before the LDAMP algorithm as a preprocessing process before compressed sensing reconstruction. The preprocessing reduces the interference of high-frequency details of the image on low-frequency information during compressed sensing reconstruction, which greatly improves the image reconstruction effect of the algorithm.

[0087] The present invention also provides an image compressed sensing reconstruction system, including a degradation module and a reconstruction module; the degradation module is used to perform degradation processing on the original image to be processed using a preset first denoising model to obtain a degraded image; the reconstruction module is used to perform compressed reconstruction processing on the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compressed sensing reconstruction result; wherein, the improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoising in the deep learning-based approximate message passing algorithm with a preset second denoising model; the preset first denoising model and the preset second denoising model are both grayscale image Gaussian noise denoising network models based on image prior modeling.

[0088] The present invention also provides an image compressed sensing reconstruction device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the image compressed sensing reconstruction method.

[0089] When the processor executes the computer program, it implements the steps of the above-mentioned image compressed sensing reconstruction method, for example: using a preset first denoising model to perform degradation processing on the original image to be processed, and obtain the degraded image;

[0090] An improved deep learning-based approximate message passing algorithm is used to compress and reconstruct the degraded image, resulting in the image compressed sensing reconstruction.

[0091] The improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; the preset first denoiser model and the preset second denoiser model are both grayscale image Gaussian noise denoiser network models based on image prior modeling.

[0092] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system. For example, a degradation module is used to perform degradation processing on the original image to be processed using a preset first denoising model to obtain a degraded image; a reconstruction module is used to perform compression reconstruction processing on the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compression-sensory reconstruction result; wherein, the improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoising in the deep learning-based approximate message passing algorithm with a preset second denoising model; the preset first denoising model and the preset second denoising model are both grayscale image Gaussian noise denoising network models based on image prior modeling.

[0093] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, the instruction segments describing the execution process of the computer program in the image compression sensing reconstruction device. For example, the computer program can be divided into a degradation module and a reconstruction module, with the specific functions of each module as follows: the degradation module is used to perform degradation processing on the original image to be processed using a preset first denoising model to obtain a degraded image; the reconstruction module is used to perform compression reconstruction processing on the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compression sensing reconstruction result; wherein, the improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoising mechanism in the deep learning-based approximate message passing algorithm with a preset second denoising model; the preset first denoising model and the preset second denoising model are both grayscale image Gaussian noise denoising network models based on image prior modeling.

[0094] The image compression sensing reconstruction device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The image compression sensing reconstruction device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are merely examples of image compression sensing reconstruction devices and do not constitute a limitation on the image compression sensing reconstruction device. It may include more components than described above, or combine certain components, or different components. For example, the image compression sensing reconstruction device may also include input / output devices, network access devices, buses, etc.

[0095] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center of the image compression sensing reconstruction device, connecting all parts of the device via various interfaces and lines.

[0096] The memory can be used to store the computer program and / or modules. The processor implements various functions of the image compression sensing reconstruction device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0097] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback or image playback). The data storage area may store data created based on the use of the phone (such as audio data or a phonebook). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMediaCards (SMC), Secure Digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0098] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the image compressed sensing reconstruction method.

[0099] If the modules / units integrated in the image compression sensing reconstruction system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0100] Based on this understanding, the present invention can implement all or part of the processes in the above-described image compressed sensing reconstruction method, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described image compressed sensing reconstruction method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0101] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0102] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

[0103] Example

[0104] This embodiment provides an image compressed sensing reconstruction method, which specifically includes the following steps:

[0105] Step 1: Construct the grayscale image Gaussian noise denoising network model based on image prior modeling; the grayscale image Gaussian noise denoising network model based on image prior modeling includes several levels of noise models; wherein, each level of noise model includes a first convolutional layer, a first multi-scale fusion module, a first supervised attention module, a second multi-scale fusion module, a channel merging module, a second supervised attention module, a second convolutional layer, an attention module, a first residual connection module, a third convolutional layer, and a second residual connection module; wherein, the previous level noise model is a preprocessing model for the next level noise model, used to fine-tune the training of the next level noise model.

[0106] The working principle of each noise model is as follows:

[0107] Step 11: Let the original clean image be denoted as x, and the input noisy grayscale image be denoted as x. n The image size remains unchanged, but the number of channels becomes 64 by performing a convolution operation through the first convolutional layer, resulting in the initial feature map F. head .

[0108] Step 12: Plot the initial feature map using F. head The feature map F is obtained by using it as input to the first multi-scale fusion module for feature extraction. m1 The feature extraction process is as follows:

[0109] Step 121: For the initial feature map F head After a 3×3 convolution layer, the number of channels is halved, and then features are extracted through a deformable attention module. The deformable attention module includes a deformable convolution layer followed by a parallel spatial attention unit (SA) and a channel attention unit (CA). The result of the deformable convolution is input to the SA and CA units through a 3×3 convolution layer. The output features of the SA and CA units are merged through channels and then subjected to residual operation with the input features of the SA and CA units through a 3×3 convolution layer. The input and output feature sizes of the deformable attention module are kept consistent.

[0110] Step 122: For the initial feature map F head After another 3×3 convolution, the number of channels is halved, and then the self-attention module is used for feature extraction. The self-attention module divides the result of the input features after a 3×3 convolution into three feature matrices according to the channels. The feature matrices are then shaped and subjected to dot product and Sigmoid normalization to obtain the output features of the self-attention module.

[0111] Step 123: Merge the output features obtained in Step 121 and Step 122, and then halve the number of channels after passing through a 3×3 convolution layer.

[0112] Step 124: Transfer the initial feature map F head As input to step 122, the resulting output feature map is downsampled using discrete wavelet transform, halving the feature size and doubling the number of channels, resulting in feature map f. t1 The initial feature map F headAs input to step 123, the obtained feature map is downsampled using discrete wavelet transform, halving the feature size and doubling the number of channels, resulting in feature map f. c1 .

[0113] Step 125, Feature Map f c1 As input to step 121, feature map f t1 As input to step 122, steps 121 and 122 are repeated twice, and the feature map f c1 Obtain the lowest level feature f cn Feature map f t1 Obtain the lowest level feature f tn .

[0114] Step 126: For the input features, after passing through a 3×3 convolutional layer, the number of channels is expanded to twice the original number. Then, upsampling is performed through the inverse discrete wavelet transform (IWT), and the image feature size becomes twice the original number, while the number of channels becomes one-quarter of the original number. Finally, feature extraction is performed through a deformable attention module.

[0115] Step 127: For the input features, after passing through a 3×3 convolutional layer, the number of channels is expanded to twice the original number. Then, upsampling is performed through the inverse discrete wavelet transform (IWT). The image feature size becomes twice the original number, and the number of channels becomes one-quarter of the original number. Finally, feature extraction is performed through a self-attention module.

[0116] Step 128: Merge the channels of the features obtained in Step 126 and Step 127, and then halve the number of channels after passing through a 3×3 convolution layer.

[0117] Step 129: Transfer the feature map f cn As input to step 126, the feature map f tn As input to step 127, steps 126 to 128 are repeated twice, and the feature map f cn Obtain the original size feature f cg Feature map f tn Obtain the original size feature f tg .

[0118] Step 1210: Transfer the feature map f cg With feature map f tg Feature extraction is performed using a 3×3 convolution layer, keeping the number of channels and feature size constant. Then, the two features are fused through channel merging to obtain the feature output F of the multi-scale feature fusion module. m1 .

[0119] Step 13, Feature Map F m1The first-stage denoised image is obtained after passing through a supervised attention module (SAM).

[0120] Step 14, Feature Map F m1 Feature map F is obtained by extracting features from the input of the second multi-scale feature fusion module. m2 Feature map F m1 With feature map F m2 After channel merging, the second-stage denoised image is obtained through the attention feature extraction module.

[0121] Step 15, Feature Map F m1 With feature map F m2 After channel merging, the image undergoes channel and spatial attention mechanisms. The residual is calculated between the input image and the original image, and then a 1×1 convolutional layer reduces the number of feature channels to 1, resulting in the final noisy image. The residual between the input image and the noisy image is subtracted to obtain the denoised image in the third stage. The channel and spatial attention mechanisms used here are the same as CA and SA in the deformable attention module, starting with a 3×3 convolutional layer and ending with channel merging of the output features of CA and SA.

[0122] Step 2: Noise refinement and stratification, dividing the noise into 42 layers; that is, 42 noise levels; wherein, the noise refinement process specifically includes:

[0123] Using the standard deviation of Gaussian noise as the grading standard, the noise range required for reconstruction in the approximate message passing algorithm is divided into 42 noise levels, each corresponding to a noise model designed at level one. The correspondence between the noise level division and the noise standard deviation is shown in Table 1.

[0124] Table 1. Correspondence between Noise Level Classification and Noise Standard Deviation

[0125]

[0126] When the noise standard deviation is 0–30, starting with a noise standard deviation of 0, this range is refined with an interval length of 1, resulting in 30 noise intervals. Within the noise standard deviation range of 30–40, it is divided into two noise intervals with an interval length of 5. Within the noise standard deviation range of 40–100, it is divided into six intervals with an interval length of 10. When the noise standard deviation is greater than 100, due to the excessive Gaussian noise introduced, refining the noise intervals has little impact on the denoising performance; therefore, the original LDAMP partitioning scheme is adopted. Noise intervals can be divided with any noise standard deviation Sigma_L as the left endpoint and any positive length λ∈R. +The step size is [Sigma]. The interval is divided into [Sigma]. L Sigma L +λ].

[0127] Step 3: Construct the training dataset; 1120 clean grayscale images are selected from the DIV2K and SIDD datasets as the real dataset, and different levels of Gaussian noise are added to the clean images to create the noise dataset. The real images and noisy images are used as a set of training data; the specific process is as follows:

[0128] Step 31: Select 800 images from DIV2K and 320 images from SIDD, and convert them to grayscale images as clean contrast images.

[0129] Step 32: Crop the image to a size of 256*256, resulting in 26K images.

[0130] Step 33: During the training process, noise is added to the clean image for different noise levels. The noise mean is 0, and the standard deviation is a random value within the standard deviation range corresponding to each noise level. The noisy image and the corresponding clean image are sent as a pair of training data into the network for training.

[0131] Step 4: Progressively train the noise model at each level.

[0132] Based on the noise levels defined in step 2, the noise model designed in step 1 is iteratively trained using the training dataset described in step 3. The specific training scheme is as follows.

[0133] Step 41: The training image size is 256×256. Each training image is normalized to the range [0,1] before being fed into the network. The batch size is set to 8. The initial learning rate is set to 0.0001. The total number of learning epochs is set to 300. After iterating the entire training dataset 120 times, the learning rate is reduced to 1 / 2 of the original. After that, the learning rate is reduced by 1 / 2 every 60 iterations of each pair of training datasets.

[0134] Step 42: Use all denoising results during training. The mean squared error (MSE) of the original clean image is used as the loss function, and the specific expression of the loss function is as follows:

[0135]

[0136] Step 43: First, train the model with the smallest noise range. After training, use the trained noise model as the preprocessing model for the next noise level noise model. Fine-tune the training for the next noise model, and so on, training all 42 noise models.

[0137] Step 5: Image preprocessing degradation

[0138] Before the original image to be processed is compressed and reconstructed, it is first degraded by a noise level denoising model to reduce the high-frequency details that are difficult to recover during the compression and reconstruction process, thereby improving the overall image quality of the compressed sensing reconstruction; wherein, the noise model is obtained by traversing the noise models corresponding to 42 noise levels.

[0139] Step 6: The trained grayscale image Gaussian noise denoising network model based on image prior modeling is used as the denoising model in the iterative reconstruction process and fed into the LDAMP framework for overall compressed sensing iterative reconstruction; the improved LDAMP framework structure is the improved deep learning-based approximate message passing algorithm, as shown in the appendix. Figure 6 As shown.

[0140] Comparative experimental design:

[0141] To verify the performance of the compressed sensing reconstruction method described in this embodiment, the following two sets of experiments were conducted:

[0142] Experiment 1: Comparison of the Gaussian denoising performance of grayscale images between the denoiser proposed in this embodiment, the denoiser used in LDAMP, and existing high-performance denoisers.

[0143] Experiment 2: After the denoiser proposed in this embodiment is incorporated into the LDAMP framework, the performance of the deep learning approximate message passing algorithm (PLDAMP) based on prior modeling is compared with that of the original LDAMP; the performance of image reconstruction quality is compared with the performance of fine-tuning the noise level design after using the denoiser proposed in this embodiment; the performance of image reconstruction quality is compared before and after image degradation using the denoiser model with fine noise interval division proposed in this embodiment.

[0144] The experimental conditions were: Python 3.7; Windows 10 system; NVIDIA GTX 1080ti GPU; CUDA 9.0; PyTorch 1.1, and Peak Signal-to-Noise Ratio (PSNR) was used as the image quality evaluation standard.

[0145] In Experiment 1, the standard test set Set12 for Gaussian denoising of grayscale images was used as the test set for the experiment. The comparison algorithms included DnCNN used in the original LDAMP, as well as the high-performance deep neural network denoisers EDSR and RCAN. The noise standard deviations for comparison were 15, 25 and 50. Among them, PMCNN is the deep neural network denoiser based on prior modeling proposed in this embodiment.

[0146] In Experiment 2, two 8-bit deep, 256×256 images of Barbara and Lena were used for testing. The test images are shown in the attached image. Figure 7 The test sampling rates were 1%, 5%, and 10%; the comparison algorithm used was the original LDAMP.

[0147] Experimental results:

[0148] Table 2 shows the image denoising quality indices of different methods on the Set12 test set in Experiment 1 at noise standard deviations of 15, 25, and 50, as detailed below:

[0149] Table 2 Quality indicators of image denoising using different methods

[0150]

[0151] Here, PMCNN represents the deep neural network based on image prior modeling in this embodiment. As can be seen from Table 2 above, under the PSNR index, the network structure of this embodiment outperforms common grayscale image denoising algorithms at all noise levels. Especially at the high noise level with a noise standard deviation of 50, it outperforms other networks by more than 0.1dB on the test set Set12 and by about 1dB compared to the DnCNN network used by LDAMP. This indicates that the deep neural network based on image prior modeling proposed in this embodiment has a significant improvement over other neural networks in Gaussian denoising of grayscale images.

[0152] Table 3 presents the image reconstruction quality metrics of the test images Barbara and Lena in Experiment 2 using the compressed sensing reconstruction algorithm based on refined prior modeling proposed in this embodiment at different sampling rates for various intermediate algorithms:

[0153] Table 3 Image reconstruction quality indicators of various intermediate algorithms at different sampling rates

[0154]

[0155] PMLDAMP is a reconstruction algorithm based on deep neural networks with image prior modeling proposed in this embodiment, replacing the denoiser in LDAMP with the one proposed in this embodiment. Fine-PMLDAMP is a reconstruction algorithm based on refined image prior modeling proposed in this embodiment. Pre-PMLDAMP is the image reconstruction result after compression and reconstruction of the image after degradation using the denoiser model with refined noise interval division proposed in this embodiment.

[0156] As attached Figure 8-9 As shown, Figure 8 and Figure 9The reconstructed images of Parrot and House at a 5% sampling rate are shown in Table 3 and Appendix 4. Figure 8-9 As can be seen, the deep neural network based on image prior modeling proposed in this embodiment can significantly improve the image compressed sensing reconstruction performance of the LDAMP algorithm, and the refined noise level distribution can further improve the performance of the reconstruction algorithm. Finally, the added refined image feature prior preprocessing process can further significantly improve the reconstruction performance of the algorithm.

[0157] The descriptions of relevant parts in the image compressed sensing reconstruction system, device, and computer-readable storage medium provided in this embodiment can be found in the detailed description of the corresponding parts in the image compressed sensing reconstruction method described in this embodiment, and will not be repeated here.

[0158] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

Claims

1. An image compressed sensing reconstruction method, characterized in that, include: Using a preset first denoising model, the original image to be processed is degraded to obtain a degraded image; An improved deep learning-based approximate message passing algorithm is used to compress and reconstruct the degraded image, resulting in the image compressed sensing reconstruction. The improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; the preset first denoiser model and the preset second denoiser model are both grayscale image Gaussian noise denoiser network models based on image prior modeling. The grayscale image Gaussian noise denoising network model based on image prior modeling includes several levels of noise models; each level of noise model includes a first convolutional layer, a first multi-scale fusion module, a first supervised attention module, a second multi-scale fusion module, a channel merging module, a second supervised attention module, a second convolutional layer, an attention module, a first residual connection module, a third convolutional layer, and a second residual connection module; wherein, the previous level noise model is a preprocessing model for the next level noise model, used to fine-tune the training of the next level noise model; The first convolutional layer is used to perform convolution operations on the noisy grayscale image to obtain the initial feature map F. head ; The first multi-scale fusion module is used to process the initial feature map F. head Feature extraction is performed to obtain feature map F. m1 ; The first supervised attention module is used to process the feature map F. m1 Feature calibration is performed, and the number of channels is changed to obtain the first-stage denoised image. The second multi-scale fusion module is used to process the feature map F m1 Feature extraction is performed to obtain feature map F. m2 ; The channel merging module is used to process the feature map F. m1 and the feature map F m2 Perform channel merging to obtain the merged feature map F. m12 ; The first supervised attention module is used to process the feature map F after channel merging. m12 Feature calibration is performed, and the number of channels is changed to obtain the second-stage denoised image. The second convolutional layer is used to process the feature map F after channel merging. m12 Feature extraction is performed to obtain feature map F. c2 ; Attention module, used for the feature map F c2 Channel and spatial attention features are extracted to obtain feature map F. csa ; The first residual connection module is used to process the initial feature map F. hea and the feature map F csa Summation is performed to obtain the feature map F. res1 ; The third convolutional layer is used to process the feature map F. res1 Feature extraction is performed to obtain feature map F. c3 ; The second residual connection module is used to connect the noisy grayscale image with the feature map F. c3 Summation is performed to obtain the third-stage denoised image.

2. The image compressed sensing reconstruction method according to claim 1, characterized in that, The training process for each level of the noise model is as follows: Construct a training dataset; wherein the training dataset includes several pairs of training data; each pair of training data includes real data and noise data, wherein the real data is a clean grayscale image selected from the DIV2K dataset and the SIDD dataset, and the noise data is Gaussian noise of a preset noise level added to the clean grayscale image; Normalize each training data point to obtain normalized training data; The normalized training data is used as input to the current noise model, and the current noise model is trained based on the target loss function; wherein, the target loss function is based on the clean grayscale image and the first-stage denoised image. Second stage denoised image and the third-stage denoised image Calculated; Once the current noise model is trained, it is used as a preprocessing model for the next level noise model, and the next level noise model is fine-tuned. This process is repeated to train all noise models.

3. The image compressed sensing reconstruction method according to claim 2, characterized in that, The target loss function is: Where MSE(*) is the mean squared error; x is a clean grayscale image; This is the first stage of denoising the image; For the second stage of denoising; This is the image for the third stage of denoising.

4. The image compressed sensing reconstruction method according to claim 1, characterized in that, The level of each noise model is adapted to the noise level of the original image to be processed; wherein, the noise level of the original image to be processed is obtained by dividing the noise range of the original image to be processed using the standard deviation of Gaussian noise as the grading standard.

5. The image compressed sensing reconstruction method according to claim 1, characterized in that, The first multi-scale fusion module and its subordinate second multi-scale fusion module are both multi-scale feature fusion modules; the multi-scale feature fusion module includes a Transformer branch, a CNN branch, a first channel merging-convolution unit, a second channel merging-convolution unit, a third channel merging-convolution unit, a fourth channel merging-convolution unit, and a fifth channel merging-convolution unit; The Transformer branch includes a first convolutional layer unit, a first self-attention module, a first discrete wavelet transform unit, a second convolutional layer unit, a second self-attention module, a second discrete wavelet transform unit, a third convolutional layer unit, a third self-attention module, a fourth convolutional layer unit, an inverse transform unit of the first discrete wavelet transform, a fourth self-attention module, a fifth convolutional layer unit, an inverse transform unit of the second discrete wavelet transform, a fifth self-attention module, a sixth convolutional layer unit, and a first output unit, connected in sequence. The CNN branch includes, in sequence, a seventh convolutional layer unit, a first deformable attention module, a third discrete wavelet transform unit, an eighth convolutional layer unit, a second deformable attention module, a fourth discrete wavelet transform unit, a ninth convolutional layer unit, a third deformable attention module, a tenth convolutional layer unit, an inverse transform unit of the third discrete wavelet transform, a fourth deformable attention module, an eleventh convolutional layer unit, an inverse transform unit of the fourth discrete wavelet transform, a fifth deformable attention module, a twelfth convolutional layer unit, and a second output unit; The outputs of the first self-attention module and the first deformable attention module are both connected to the input of the first channel merging-convolution unit, and the output of the first channel merging-convolution unit is connected to the input of the third discrete wavelet transform unit. The outputs of the second self-attention module and the second deformable attention module are both connected to the input of the second channel merging-convolution unit, and the output of the second channel merging-convolution unit is connected to the input of the fourth discrete wavelet transform unit. The outputs of the third self-attention module and the third deformable attention module are both connected to the input of the third channel merging-convolution unit, and the output of the third channel merging-convolution unit is connected to the input of the tenth convolutional layer unit. The outputs of the fourth self-attention module and the fourth deformable attention module are both connected to the input of the fourth channel merging-convolution unit, and the output of the fourth channel merging-convolution unit is connected to the input of the eleventh convolutional layer unit. The outputs of the fifth self-attention module and the fifth deformable attention module are both connected to the input of the fifth channel merging-convolution unit, and the output of the fifth channel merging-convolution unit is connected to the input of the twelfth convolutional layer unit.

6. The image compressed sensing reconstruction method according to claim 1, characterized in that, The process of compressing and reconstructing the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compressed sensing reconstruction result is as follows: When performing the iterative reconstruction calculation at level t, the divergence of the noise model at level t. for: The deviation z between the calculated reconstruction result of the t-th layer iterative reconstruction calculation and the actual measurement value. t for: The estimated standard deviation σ between the calculated reconstruction result and the actual measured value in the t-th layer iterative reconstruction calculation. t for: The iterative reconstruction calculation at layer t produces data b of a standard normal distribution with size [1, N]: b = rand(1, N) Where x is the original image to be processed, x D Let y be the degraded image and x be the degraded image. D The measured values ​​after the measurement matrix is ​​applied; M is the number of measured values; N is the total number of pixels in the image, x t Let |x|| be the reconstructed data vector of layer t; ∞ Let b be the infinite norm of the original image x to be processed. H Let A be the transpose of data b, and let A be the measurement matrix. H To measure the transpose of the matrix, This represents a noise model.

7. An image compressed sensing reconstruction system, characterized in that, include The degradation module is used to perform degradation processing on the original image to be processed using a preset first denoising model to obtain a degraded image; The reconstruction module is used to perform compressed reconstruction processing on the degraded image using an improved deep learning-based approximate message passing algorithm to obtain the image compressed sensing reconstruction result. The improved deep learning-based approximate message passing algorithm is an algorithm that replaces the denoiser in the deep learning-based approximate message passing algorithm with a preset second denoiser model; the preset first denoiser model and the preset second denoiser model are both grayscale image Gaussian noise denoiser network models based on image prior modeling. The grayscale image Gaussian noise denoising network model based on image prior modeling includes several levels of noise models; each level of noise model includes a first convolutional layer, a first multi-scale fusion module, a first supervised attention module, a second multi-scale fusion module, a channel merging module, a second supervised attention module, a second convolutional layer, an attention module, a first residual connection module, a third convolutional layer, and a second residual connection module; wherein, the previous level noise model is a preprocessing model for the next level noise model, used to fine-tune the training of the next level noise model; The first convolutional layer is used to perform convolution operations on the noisy grayscale image to obtain the initial feature map F. head ; The first multi-scale fusion module is used to process the initial feature map F. head Feature extraction is performed to obtain feature map F. m1 ; The first supervised attention module is used to process the feature map F. m1 Feature calibration is performed, and the number of channels is changed to obtain the first-stage denoised image. The second multi-scale fusion module is used to process the feature map F m1 Feature extraction is performed to obtain feature map F. m2 ; The channel merging module is used to process the feature map F. m1 and the feature map F m2 Perform channel merging to obtain the merged feature map F. m12 ; The first supervised attention module is used to process the feature map F after channel merging. m12 Feature calibration is performed, and the number of channels is changed to obtain the second-stage denoised image. The second convolutional layer is used to process the feature map F after channel merging. m12 Feature extraction is performed to obtain feature map F. c2 ; Attention module, used for the feature map F c2 Channel and spatial attention features are extracted to obtain feature map F. csa ; The first residual connection module is used to process the initial feature map F. head and the feature map F csa Summation is performed to obtain the feature map F. res1 ; The third convolutional layer is used to process the feature map F. res1 Feature extraction is performed to obtain feature map F. c3 ; The second residual connection module is used to connect the noisy grayscale image with the feature map F. c3 Summation is performed to obtain the third-stage denoised image.

8. An image compression sensing reconstruction device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the image compressed sensing reconstruction method as described in any one of claims 1-6 when executing the computer program.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the image compressed sensing reconstruction method as described in any one of claims 1-6.

Images