Image compression sensing reconstruction method and system based on PIPO-Net

Abstract

The present invention discloses a method and system for image compression sensing reconstruction based on PIPO‑Net. The present invention obtains a two-dimensional image; preprocesses the two-dimensional image; inputs the preprocessed two-dimensional image into a penalty-based independent parameter optimization deep expansion network PIPO‑Net to obtain a compression sensing measurement value of the two-dimensional image, and finally outputs a reconstructed image. PIPO‑Net can obtain the average value of the original image and the image after de-averaging by a de-averaging sampling method. The sampling values ​​of the image after de-averaging are used for training, thereby improving the performance and stability of the network and better capturing the features of the image. In order to solve the problem of information loss, the present invention designs a high-frequency information supplement block and uses a wavelet high-frequency loss function for training to better retain the details and texture information of the image.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and image compression sensing, and specifically relates to an image compression sensing reconstruction method and a system thereof based on PIPO-Net. Background Art

[0002] Compressed sensing is one of the emerging tasks in the field of computer vision. It aims to achieve data compression and transmission more efficiently by effectively capturing and utilizing the structural features and redundant information of signals such as images and videos. Compared with traditional compression methods based on signal sampling, compressed sensing methods can significantly reduce the sampling rate and storage space while maintaining high-quality reconstruction, thereby saving computing and storage costs and improving resource utilization efficiency. It has a wide range of application value and can be applied to industrial design, medical imaging, sensor machinery, public safety and other fields. However, it is time-consuming and labor-intensive to achieve compressed sensing image reconstruction through traditional algorithms. Therefore, image generation based on model algorithms has become a major research hotspot in the field of computer vision due to its advantages of automation and low cost.

[0003] Traditional iterative solutions have advantages such as strong interpretability, but they also have obvious disadvantages, such as requiring more hardware resources and being time-consuming. Although the solution using traditional deep learning models can be reconstructed at a faster speed and with good results, it lacks the interpretability and structural simplicity of traditional iterative solutions. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for image compression sensing reconstruction based on PIPO-Net in view of the deficiencies of the prior art, that is, to design a deep unfolding network architecture (PIPO-Net) for image compression sensing reconstruction. All parameters of PIPO-Net are self-learning, thereby improving the adaptability and performance of the network. PIPO-Net has good interpretability and makes full use of the advantages of compression sensing reconstruction methods based on optimized iterative algorithms and classical deep networks, achieving more advanced computational efficiency.

[0005] In a first aspect, the present invention provides an image compression sensing reconstruction method based on PIPO-Net, the method comprising the following steps:

[0006] Step 1: Obtain a two-dimensional image;

[0007] Step 2: preprocessing the two-dimensional image; the preprocessing method includes two-dimensional image scaling and cropping;

[0008] Step 3: Input the preprocessed two-dimensional image into the penalty-based independent parameter optimization deep unfolding network PIPO-Net to obtain the compressed sensing measurement value of the two-dimensional image, and finally output the reconstructed image;

[0009] The penalty-based independent parameter optimization deep unfolding network PIPO-Net includes a sampling phase and a reconstruction phase;

[0010] The sampling stage is to rearrange each input image into a plurality of vectorized image block data of size n×1, and use a de-meaning sampling strategy to obtain a de-meaning measurement value and an image block pixel mean, thereby realizing compressed sampling of the vectorized image block data;

[0011] The reconstruction stage includes an initial reconstruction module IRM and a depth reconstruction layer;

[0012] The initial reconstruction module IRM performs initial reconstruction on the de-averaged measurement values ​​and then transmits them to the depth reconstruction layer;

[0013] The depth reconstruction layer includes K serially connected depth reconstruction modules DRM.

[0014] The initial reconstruction module consists of a fully connected layer, which converts the input sampling values ​​into an initial reconstruction image.

[0015] The depth reconstruction module DRM includes an update unit and a high-frequency information supplement block HFC; the update unit includes a serially connected update auxiliary block, a multiplier update block, and an image update block;

[0016] During the training process of PIPO-Net, the loss value of the reconstructed image output by the network and the target image is calculated through the spatial domain loss function, and the loss value of the output image of the high-frequency information supplement block of each deep reconstruction module DRM and the target model is calculated using the wavelet loss function; by fusing the back propagation of the two loss values, the network iteration parameters continuously reduce the loss value to complete the training.

[0017] In a second aspect, the present invention provides an image compression sensing reconstruction system, comprising:

[0018] A data acquisition module is used to acquire two-dimensional images and construct a training set;

[0019] A data preprocessing module, used for preprocessing the two-dimensional image;

[0020] The image compressed sensing reconstruction module is used to input the preprocessed two-dimensional image into the penalty-based independent parameter optimization deep expansion network PIPO-Net to obtain the compressed sensing measurement value of the two-dimensional image and finally output the reconstructed image.

[0021] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

[0022] In a fourth aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

[0023] The beneficial effects of the present invention are as follows:

[0024] 1. The PIPO-Net architecture based on independent parameter optimization of penalty terms proposed in the present invention makes full use of the prior information and mean information of the original image, combines deep learning methods with traditional optimization methods, and can solve the compressed sensing problem under different compression rates. The method is easy to implement, the data preprocessing operation is simple, and it has better accuracy.

[0025] 2. All parameters in the penalty-based independent parameter optimization deep expansion network PIPO-Net proposed in the present invention are self-learning, which can optimize the hyperparameters of the algorithm from the training data, greatly reducing the tedious process of manual operation.

[0026] 3. The present invention uses a hybrid loss function to communicate and constrain the image generation model in the wavelet domain and the image domain to train more accurately, thereby improving the quality of high-frequency information generated by the image and solving the problem of block effect noise in image reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a detailed structural diagram of the deep unfolding network architecture of image compressed sensing reconstruction in the present invention.

[0028] Figure 2 The Net z And the detailed structure diagram of the high-frequency information supplement block.

[0029] Figure 3 The images of the ship reconstructed by all seven methods when the CS compression ratio is 25%.

[0030] Figure 4 The images of Barbara reconstructed by all seven methods when the CS compression ratio is 10%. DETAILED DESCRIPTION

[0031] The present invention is further analyzed below in conjunction with specific embodiments.

[0032] A deep unfolding network architecture for image compressed sensing reconstruction, including the following:

[0033] A data acquisition module is used to acquire two-dimensional images and construct a training set;

[0034] A data preprocessing module, used for preprocessing the two-dimensional image;

[0035] The image compression sensing reconstruction module is used to input the preprocessed two-dimensional image into the penalty-based independent parameter optimization deep expansion network PIPO-Net to obtain the compression sensing measurement value of the two-dimensional image and finally output the reconstructed image;

[0036] The penalty-based independent parameter optimization deep unfolding network PIPO-Net includes a sampling stage and a reconstruction stage; the sampling stage is used to perform compressed sensing sampling on the original image, and the compressed sampling of the original image block is achieved through the mean removal sampling strategy. The output of this sampling stage includes the mean removal measurement value and the image block pixel mean; the reconstruction stage is used to restore the mean removal measurement value to the original image, and the reconstruction stage includes an initial reconstruction module and K deep reconstruction modules. In the initial reconstruction module, the mean removal measurement value is used to achieve initial reconstruction. By transmitting through K deep reconstruction modules in sequence, high-precision reconstruction can be obtained; in order to eliminate blocking artifacts and restore lost high-frequency information, an additional high-frequency information supplement block is added to the end of each deep reconstruction module, and each basic image block can be restored using the basic image signal and the patch pixel mean; the restoration proposed by the present invention fully utilizes the advantages of the compressed sensing reconstruction method based on the optimized iterative algorithm and the classical deep network, and achieves more advanced computational efficiency.

[0037] A deep unfolding network architecture for image compressed sensing reconstruction, including the following contents:

[0038] Step 1: Obtain the required data set, the data source is the BSD500 public data set.

[0039] Step 2: Preprocess the two-dimensional images in the data set. The preprocessing methods include two-dimensional image scaling and cropping.

[0040] Step 3: Build and train the neural network.

[0041] 3-1. Building a complete reconstruction network - PIPO-Net based on independent parameter optimization of penalty terms

[0042] The penalty-based independent parameter optimization deep unfolding network PIPO-Net includes a sampling phase and a reconstruction phase, such as Figure 1 shown.

[0043] In the sampling stage, each input image is first rearranged into vectorized image block data of size n×1. The de-meaning sampling strategy is used to obtain the de-meaning measurement value and the image block pixel mean, thereby realizing the compressed sampling of the training vector.

[0044] In this embodiment, the specific process of the de-averaging sampling strategy is as follows:

[0045] Let i be the number of iterations, and let the quantized image block data be The pixel mean of the image block is

[0046] According to the arithmetic mean of the image block, it is expanded into a vector x with the same dimension as the original vectorized image block i :

[0047]

[0048] Perform de-meaning on the vectorized image block to obtain the de-meaning vector x i :

[0049]

[0050] The de-meaning sampling process to be achieved by compressed sensing can be formulated as:

[0051] y i =Ax i

[0052] in represents the original sampling matrix;

[0053] Since x before sampling * i Unknown, so it is impossible to obtain the image data after removing the mean; to solve this problem, construct a new augmented sampling matrix:

[0054]

[0055] That is, a row of n all-1 sampling vectors is expanded after the original sampling matrix A; based on this, the actual sampling process is recorded as:

[0056]

[0057] in y i,m+1 represents the total number of pixels in the image block, then The pixel mean of the image block can be obtained

[0058] The actual sample value y * i Calculate the mean measured value y i :

[0059]

[0060] In summary, by introducing the augmented sampling matrix, the mean-removing compressed sampling of the image is achieved.

[0061] The reconstruction phase includes an initial reconstruction module and K series-connected deep reconstruction modules DRM. In the initial reconstruction module, the mean measurement values ​​are removed to achieve initial reconstruction. By transmitting through K deep reconstruction modules in sequence, a high-precision reconstruction can be obtained. In order to eliminate blocking artifacts and restore the lost high-frequency information, an additional high-frequency information supplement block is added to the end of each deep reconstruction module. Using the basic image signal and the patch pixel mean, each basic image block can be restored. The image is then quickly reassembled into the entire image and enters the high-frequency information supplement block. Then, the output image of the high-frequency information supplement block is divided into blocks again and output as the next deep reconstruction module.

[0062] In this embodiment, the depth reconstruction module DRM includes an update unit and a high-frequency information supplement block HFC; the update unit includes a serially connected update auxiliary block, a multiplier update block, and an image update block;

[0063] The specific implementation process of updating the auxiliary block is:

[0064] Auxiliary variables The update mode is:

[0065]

[0066] in represents the output of the previous deep reconstruction module, Represents the output of the initial reconstruction module IRM; represents the penalty parameter that can be learned by the current deep reconstruction module, which is independent of the penalty parameters of other DRMs; i-1 (k) Represents the multiplier vector updated by the i-1th iteration of the current depth reconstruction module; Net z Represents the subnetwork for updating the auxiliary variable z, which includes three parts: feature extraction, feature enhancement and feature aggregation; the feature extraction part includes convolution filter CONV1, batch normalization (BN) and rectified linear unit (ReLU), which aims to extract the features of the input data; the feature enhancement part includes the first residual block RES1, the rectified linear unit (ReLU), and the second residual block RES2, which further extracts and enhances the features extracted by the feature extraction part; the feature aggregation part includes batch normalization (BN), rectified linear unit (ReLU) and convolution layer CONV2, which integrates the features enhanced by the feature enhancement part and extracts deeper features; the input and output are fused together through skip connections, which can promote the convergence of the subnetwork.

[0067] The multiplier update block updates the auxiliary variables obtained by the auxiliary update block. The multiplier vector is updated, which is expressed as:

[0068]

[0069] Each depth reconstruction module pre-allocates the initial multiplier vector λ0 in the memory space (k) are all 0, and the i-th iteration update is Deposit The storage space where the parameter;

[0070] The image update block updates the auxiliary variables obtained by the auxiliary update block and the multiplier update block. Multiplier Vector Update the basic image signal to obtain the output of the update unit It is expressed as:

[0071]

[0072] Where A is the sampling matrix, y i is the mean-removed sampling value of the i-th iteration, I n represents the identity matrix of size n×n;

[0073] In the image reconstruction process, it is necessary to consider noise suppression and the supplementation of missing high-frequency information. Combined with the wavelet loss function, a high-frequency information supplement block is designed, and a data-driven method is used for learning and optimization. By adding the high-frequency information learned by the high-frequency information supplement block to the output of each layer, the reconstruction result is closer to the original signal. Figure 2 The high frequency information supplement block HFC is a CNN network, and its structure is similar to that of Net z Consistent, including feature extraction, feature enhancement and feature aggregation; update the output of the unit Add the pixel mean of the image block After that, it is sent to the high-frequency information supplement block HFC, and the features are learned using deep supervision. Then, the result after deep supervision learning is compared with the pixel mean x of the image block. i * Subtract to get the final output x of the depth reconstruction module i (k) ;

[0074] 3-2. Model training

[0075] The reconstructed image data is input into the deep learning network recorded in step 3-1 for training. The training will not only calculate the loss value of the reconstructed model output by the network and the target model through the spatial domain loss function, but also use the wavelet loss function to calculate the loss value of the output image of the high-frequency information supplement block of each deep reconstruction module and the target model. Through the back propagation of the fused loss value, the network iteration parameters continuously reduce the loss value to complete the training. The present invention uses the fused loss function to make the loss value calculation more comprehensive.

[0076] The fusion loss value, input image X, model reconstructed image Xi. In the spatial domain, the input image X and the model reconstructed image Xi are directly subjected to the L2 loss function to obtain the loss value. In the wavelet domain space, the wavelet transform is applied to decompose the two into sub-band graphs [A, V, H, D], where A corresponds to low-frequency information, V represents vertical high-frequency information, H represents horizontal high-frequency information, and D represents diagonal high-frequency information. The calculation objects of the wavelet loss function are X and Xi, and the four sub-band graphs representing the high-frequency and low-frequency information of the two are spliced, and the L2 loss is calculated using the spliced ​​self-contained graph. Calculating the loss separately on different domains of the image replaces the traditional calculation of the loss on the image pixel value to improve the quality of the image generated by the trained image generation model.

[0077] Step 4: According to the network model built and trained in step 3, the compressed sensing measurement value of a single two-dimensional image is input, and finally the reconstructed image is output.

[0078] The workflow of the basic reconstruction network is as follows: In the basic reconstruction network, first input a 264*264 size image, select a set compression ratio, perform sampling in the sampling stage, obtain the mean-removed measurement value and the image block pixel mean, and send them to the reconstruction stage; in the reconstruction stage, first enter the initial reconstruction module, convert the mean-removed measurement value into an initial reconstruction image through a fully connected layer, and realize initial reconstruction. Then, the initial reconstruction image is transmitted through K deep reconstruction modules in sequence to obtain a high-precision reconstruction, and finally re-output a 264*264 size image;

[0079] The fusion loss function in step 3-2 model training is shown in formula (9), which is the sum of the fusion loss space loss function and the wavelet loss function. In order to reduce the gap between the network reconstructed image and the input image, it is also necessary to learn the high-frequency information of the reconstructed image at each layer stage. The spatial domain loss function L of the present invention is MSE Use the average normalized root mean square error (MSE):

[0080]

[0081] Numel(·) counts the number of pixels in an image, which is the output image corresponding to the network. ||·||F is the Frobenius norm, N represents the number of complete two-dimensional images selected for the jth training, X j represents the complete original input 2D image, Represents the image reconstructed by the deep unfolding network PIPO-Net based on independent parameter optimization of penalty terms;

[0082] The following wavelet loss function:

[0083]

[0084] Among them, WT(·) is the Haar wavelet decomposition operation, represents the high frequency information supplement block HFC of the k-th deep reconstruction module DRM, Represents the output of the update unit of the k-th depth reconstruction module DRM.

[0085] Combining these two loss functions, the total loss function is:

[0086] L total =L mse +γL wt (9)

[0087] Where γ is the regularization parameter, which is set to 0.01. When using the wavelet loss function, in order to keep the brightness of the reconstructed image consistent with that of the input image, the network will add the mean of the input image back to the reconstructed image. After using the loss function, before inputting the lower layer for reconstruction, the network automatically subtracts the mean of the original image from the output and sends it to the next layer.

[0088] In order to verify the effectiveness of the present invention, a variety of existing reconstruction methods including ReconNet, CSNet, ISTA-Net, BCSnet, iPiano-Net, and AMP-Net are used. Among them, ReconNet and CSNet are compressed sensing reconstruction methods based on deep networks, and BCSnet, iPiano-Net, ISTA-Net, and AMP-Net are compressed sensing reconstruction methods based on deep expansion. The experiment selects 5 different categories of compression rates based on the SET11 data set for testing, and uses PSNR value and SSIM value as evaluation indicators.

[0089] The formula for PSNR is as follows:

[0090]

[0091] Among them, MAX is the maximum possible value of the original image or video pixel, and MSE is the mean square error. The unit of PSNR is decibel (dB). Generally speaking, the higher the PSNR value, the better the quality of the image or video.

[0092] The calculation of SSIM is based on brightness, contrast and structural similarity measures within a local window, as well as a constant C for stability. The calculation formula of SSIM is as follows:

[0093]

[0094] Among them, x and y represent the original image and the processed image respectively, μ x and μ y Represent their local brightness means, σx and σ y Represent their local variances, σ xy represents their local covariance. C1 and C2 are constants used to avoid the denominator being zero. The result of SSIM is between -1 and 1. The closer the value is to 1, the higher the structural similarity between images, that is, the better the image quality. Compared with PSNR, SSIM can better reflect the human eye's perception of image quality. Therefore, in image processing tasks, SSIM is generally considered to be more effective than PSNR.

[0095] Compared with other methods, the network proposed in the present invention has achieved overall superior performance in reconstructing images. The PIPO-Net proposed in the present invention can achieve good reconstruction accuracy under different compression ratios, achieve the best results under all compression ratios, and reconstruct clearer images. It can be noted that with the MSS strategy and HFC, the effect of PIPO-Net is better than other algorithms. Compared with AMP-Net, the network model of the present invention has a significant improvement in each compression rate on the Set11 dataset. At low compression rates, that is, at compression rates of 1% and 10%, it is 1.47dB and 0.42dB higher than AMP-Net respectively. In addition, at a compression rate of 50%, the network model of the present invention reached 40.43dB. At the same time, Table 3-1 shows the average running time (in seconds) of various algorithms to reconstruct images, which clearly shows that PIPO-Net produces better reconstruction results while maintaining a faster reconstruction speed. Figure 3 The images of the ship reconstructed using all seven methods are shown when the CS compression ratio is 25%.

[0096] Table 3-1 Comparison in Set11 dataset

[0097]

[0098] In order to verify the robustness and generalization of the proposed method on different datasets, experiments were also conducted on the BSD68 dataset. The experimental results on the BSD68 dataset are shown in Table 3-2.

[0099] Table 3-2 Comparison in the BSD68 dataset

[0100]

[0101] According to the experimental results on the BSD68 dataset, the best results are marked in bold. Compared with other methods, the network proposed in the present invention also performs better at different compression rates. Compared with AMP-Net on the BSD68 dataset, except for the compression rate of 50%, both PSNR and SSIM are the best. This is mainly due to our independent parameter optimization model, which enables the network to find the optimal solution more flexibly. Compared with other methods, the method proposed in the present invention shows better performance and better generalization ability on different datasets. PIPO-Net performs best because each layer of PIPO-Net can be independently regarded as a solution process for a compressed sensing problem, and the reconstruction results of each layer are independent. In addition, all parameters in PIPO-Net are self-learned, and PIPO-Net can optimize the hyperparameters of the algorithm from the training data. In addition, the deep supervision method uses the high-frequency loss function to supervise the learning of high-frequency information, which not only reduces the block effect, but also enhances the quality of the reconstructed image. Especially at low compression ratios of 10% and 1%, the advantage of PIPO-Net is more significant. This can be explained by the fact that in the case of low compression ratios, the key information of the original signal may be severely lost due to the sampling process. Without additional auxiliary information, the performance of most networks is very poor, as shown in Table 1. For the PIPO-Net proposed in the present invention, due to the use of the MSS strategy, the prior pixel average information is used as additional assistance to improve the reconstruction performance. At the same time, Table 1 shows the parameter amounts of various reconstruction algorithms. Although the parameter amount is slightly higher than the best result, it is still a relatively good performance. This clearly shows that the PIPO-Net proposed in the present invention can produce better reconstruction results faster while maintaining a lower parameter amount and can be more easily deployed on the server. Figure 4 Visuals of the PSNR / SSIM results for the reconstructed Barbara image with a compression ratio of 10% are shown. Examining the magnified portion of the reconstructed image, the proposed PIPO-Net recovers more texture details and the results appear to be more realistic. Especially for the Babrar case with a compression ratio of 10%, in addition to the gains obtained by the MSS strategy, the HFC block effectively compensates for the missing high-frequency information by absorbing the wavelet loss function, eliminating the undesirable blocking artifacts. These results highlight the superior performance of PIPO-Net.

Claims

1. A method for image compression sensing reconstruction based on PIPO-Net, the method comprising the following steps: Step 1: Obtain a two-dimensional image; Step 2: preprocessing the two-dimensional image; Step 3: Input the preprocessed two-dimensional image into the penalty-based independent parameter optimization deep unfolding network PIPO-Net to obtain the compressed sensing measurement value of the two-dimensional image, and finally output the reconstructed image; It is characterized in that The penalty-based independent parameter optimization deep unfolding network PIPO-Net includes a sampling phase and a reconstruction phase; The sampling stage is to rearrange each input image into a plurality of vectorized image block data of size n×1, and use a de-meaning sampling strategy to obtain a de-meaning measurement value and an image block pixel mean, thereby realizing compressed sampling of the vectorized image block data; The reconstruction stage includes an initial reconstruction module IRM and a depth reconstruction layer; The initial reconstruction module IRM performs initial reconstruction on the de-averaged measurement values ​​and then transmits them to the depth reconstruction layer; The depth reconstruction layer includes K series-connected depth reconstruction modules DRM; The depth reconstruction module DRM includes an update unit and a high-frequency information supplement block HFC; the update unit includes a serially connected update auxiliary block, a multiplier update block, and an image update block; The specific implementation process of updating the auxiliary block is: Auxiliary variables The update mode is: where x i (k-1) represents the output of the previous deep reconstruction module, x i (0) Represents the output of the initial reconstruction module IRM; represents the penalty parameter that can be learned by the current deep reconstruction module, which is independent of the penalty parameters of other DRMs; i-1 (k) Represents the multiplier vector updated by the i-1th iteration of the current depth reconstruction module; Net z Represents the subnetwork for updating the auxiliary variable z, which includes three parts: feature extraction, feature enhancement and feature aggregation; the feature extraction part includes convolution filter CONV1, batch normalization (BN) and rectified linear unit (ReLU); the feature enhancement part includes the first residual block RES1, rectified linear unit (ReLU), and the second residual block RES2, which further extracts and enhances the features extracted by the feature extraction part; the feature aggregation part includes batch normalization (BN), rectified linear unit (ReLU) and convolution layer CONV2, which integrates the features enhanced by the feature enhancement part to extract deeper features; The multiplier update block updates the auxiliary variables obtained by the auxiliary update block. The multiplier vector is updated, which is expressed as: Each depth reconstruction module pre-allocates the initial multiplier vector λ0 in the memory space (k) are all 0, and the updated value in the i-th iteration is Deposit The storage space where the parameter; The image update block updates the auxiliary variables obtained by the auxiliary update block and the multiplier update block. Multiplier vector λ i (k) , update the basic image signal and get the output of the update unit It is expressed as: where y i is the mean-removed sampling value of the i-th iteration, I n represents the identity matrix of size n×n; The high frequency information supplement block HFC is a CNN network, and its structure is similar to that of Net z Consistent, including feature extraction, feature enhancement and feature aggregation; update the output of the unit Add the pixel mean of the image block After that, it is sent to the high-frequency information supplement block HFC, and the features are learned using deep supervision. Then, the result after deep supervision learning is compared with the pixel mean of the image block. Subtract to get the final output x of the depth reconstruction module i (k) .

2. The method according to claim 1, characterized in that: The specific process of the de-meaning sampling strategy is as follows: Let i be the number of iterations, and set the quantized image block data x i * =[x i1 ,x i2 ,...,x in ] T , where the pixel mean of the image block is According to the arithmetic mean of the image block, it is expanded into a vector with the same dimension as the original vectorized image block Perform de-meaning on the vectorized image block to obtain the de-meaning vector x i : The de-meaning sampling process to be achieved by compressed sensing can be formulated as: y i =Ax i (6) Where A represents the original sampling matrix; Since x before sampling * i Unknown, so it is impossible to obtain the image data after removing the mean; to solve this problem, construct a new augmented sampling matrix: That is, a row of n all-1 sampling vectors is expanded after the original sampling matrix A; based on this, the actual sampling process is recorded as: in y i,m+1 represents the total number of pixels in the image block, then Get the pixel mean of the image block The actual sample value y * i Calculate the mean measurement value y i : 。 3. The method according to claim 1, characterized in that: The initial reconstruction module consists of a fully connected layer, which converts the input sampling values ​​into an initial reconstruction image.

4. The method according to claim 1, characterized in that: During the training process of the deep unfolding network PIPO-Net with independent parameter optimization based on penalty terms, the loss values ​​of the reconstructed image output by the network and the target image are calculated by the spatial domain loss function, and the loss values ​​of the output image and the target model of the high-frequency information supplement block of each deep reconstruction module DRM are calculated by the wavelet loss function; by fusing the back propagation of the two loss values, the network iteration parameters continuously reduce the loss value to complete the training.

5. The method according to claim 4, characterized in that: The total loss function of the deep unfolding network PIPO-Net based on independent parameter optimization of penalty terms is expressed as: THE total =L MSE +γL WT (10) Where γ represents the weight; Spatial domain loss function L MSE Use the average normalized root mean square error (MSE): Where numel(·) represents the image pixel counting function, ||·||F is the Frobenius norm, N represents the number of complete two-dimensional images selected for the i-th training, and X j represents the complete original input 2D image, Represents the image reconstructed by the deep unfolding network PIPO-Net based on independent parameter optimization of penalty terms; Wavelet loss function L WT express: Where WT(·) is the Haar wavelet decomposition operation, represents the high frequency information supplement block HFC of the k-th deep reconstruction module DRM, Represents the output of the update unit of the k-th depth reconstruction module DRM.

6. An image compression sensing reconstruction system implementing the method according to any one of claims 1 to 5, characterized in that include: A data acquisition module is used to acquire two-dimensional images and construct a training set; A data preprocessing module, used for preprocessing the two-dimensional image; The image compressed sensing reconstruction module is used to input the preprocessed two-dimensional image into the penalty-based independent parameter optimization deep expansion network PIPO-Net to obtain the compressed sensing measurement value of the two-dimensional image and finally output the reconstructed image.

7. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 5.

8. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 5 is implemented.

Images