An ISTA model-driven EALISTA compressed sensing reconstruction algorithm and system

Abstract

The present invention relates to the field of intelligent communications, and in particular to an EALISTA compressed sensing reconstruction method driven by an ISTA model. The EALISTA algorithm uses a traditional iterative shrinkage threshold algorithm as a model, expands the traditional iterative shrinkage threshold algorithm, generates an EALISTA neural network, and pre-calculates an analysis matrix through a known sensor matrix, thereby simplifying the training structure. The network adopts a quasi-cyclic neural network structure, and each layer repeatedly inputs a sensor signal. The activation function is replaced by a threshold function according to the reconstruction model, thereby increasing interpretability and optimizing recovery performance, making the recovered sparse signal more accurate. This method passes through a multi-layer neural network, and the number of nodes in each layer of the neural network is consistent with the dimension of the signal to be recovered. It can achieve excellent recovery performance with a weight matrix of smaller dimension, and each layer of the network corresponds to each iteration in the model, which well combines the excellent recovery performance of the traditional iterative shrinkage threshold algorithm and reflects the excellent performance of the model-driven neural network.

Description

Technical Field

[0001] The present invention relates to the field of intelligent communications, and in particular to an EALISTA compressed sensing reconstruction method and system driven by an ISTA model. In large-scale machine-type communication systems, a large number of technologies are applied to sparse signals for communication transmission. This model-driven EALISTA compressed sensing reconstruction algorithm can effectively improve recovery performance. Background Art

[0002] The Nyquist sampling theorem is a well-known sampling theorem in the communications field. Its purpose is to convert analog signals in nature into digital signals that can be processed by computers while fully preserving the original signal's information. The Nyquist sampling theorem requires that the sampling frequency must be greater than or equal to twice the highest frequency in the signal's frequency domain.

[0003] This law means that in order to recover a signal that perfectly approximates the original analog signal, a large number of samples need to be collected, which is very limiting given the large number of devices and data transmission in the future.

[0004] Therefore, in 2006, Emmanuel Candès, David Donoho, and Terence Tao jointly proposed a novel data sampling compression method called compressed sensing, which achieves efficient sampling through synchronized data acquisition and signal compression. The theory holds that if the signal is sparse, it can be reconstructed and recovered from sampling points far fewer than those required by the sampling theorem.

[0005] In recent years, the rapid development of deep learning has opened up a new avenue for researchers to develop compressed sensing reconstruction algorithms using non-human-designed models. These algorithms do not rely on any prior knowledge of the signal; instead, simply by providing the neural network with sufficient training data, the algorithm can autonomously learn how to best exploit the data structure to accelerate the reconstruction process.

[0006] Methods combining deep learning with compressed sensing fall into two main categories: purely data-driven deep learning compressed sensing reconstruction algorithms and model-driven deep learning compressed sensing reconstruction algorithms. Model-driven methods are more interpretable and generally use a much smaller dimensionality of training weights than data-driven algorithms. The most representative model-driven deep learning compressed sensing reconstruction algorithm is the Learned Iterative Shrinkage Algorithm (LISTA), proposed by Karol Gregor and Yann LeCun in 2010. Building on this, Jialin Liu et al. proposed the Analytical LISTA (ALISTA), which achieves even better performance by reducing the weight matrix.

[0007] However, the ALISTA neural network still has its shortcomings: 1. Each layer of the network is trained only on independent scalars, causing network performance to fluctuate and poor convergence. 2. It is not sensitive enough to the position of the support set of sparse signals.

[0008] So we proposed a new neural network model to achieve better performance. Summary of the Invention

[0009] In response to the above problems, the present invention proposes an EALISTA compressed sensing reconstruction method and system driven by the ISTA model to improve the compressed sensing reconstruction, i.e., sparse signal recovery performance, and improve the accuracy of recovering the sparse signal support set. The EALISTA (Enhanced analytic learned iterative shrinkage threshold algorithm) algorithm is based on the traditional iterative shrinkage threshold algorithm (ISTA), expands the traditional iterative shrinkage threshold algorithm, and generates the EALISTA neural network. This method pre-calculates the analysis matrix through the known sensing matrix, thereby simplifying the training structure.

[0010] The present invention provides the following technical solution: an ISTA model-driven EALISTA compressed sensing reconstruction method, comprising the following steps:

[0011] Step 1: Based on the principle of compressed sensing, a column-normalized Gaussian matrix is ​​randomly generated as the sensing matrix, and the corresponding analysis matrix is ​​calculated based on the sensing matrix;

[0012] Step 2: Based on the sensing matrix generated in step 1, randomly generate a sparse signal that meets the requirements, and pass the sparse signal through the sensing matrix to generate a sensing signal;

[0013] Step 3: Use the sparse signal generated in step 2 as the label and the sensor signal generated in step 2 as the training data to build a data set;

[0014] Step 4: Build an EALISTA neural network based on the data set described in step 3. The network structure is a quasi-cyclic neural network structure. Each layer of the network repeatedly inputs the training data generated in step 3, and each layer of the network corresponds to an iteration of the traditional iterative shrinkage threshold algorithm. Therefore, the output result of each layer of the network can be regarded as a reconstruction of the sparse signal. The output of the upper layer of the network is passed to the lower layer of the network as iterative data and calculated together with the input each time. The activation function of each layer of the network is replaced by the threshold function of the iterative shrinkage threshold algorithm.

[0015] Step 5: Based on the network described in step 4, the error between the restored signal output by the last layer of neural network and the label is used as the loss function, and the network weights are trained using the back propagation method;

[0016] Step 6: Based on step 5, the loss function uses the mean square error between the restored signal and the label, and the optimizer uses the gradient descent optimizer, which is trained based on backpropagation.

[0017] Step 7: Based on step 5, a new loss function is used to improve the accuracy of the restored signal support set. The loss function uses a custom error function of the restored signal and the label. The optimizer uses a gradient descent optimizer and is trained based on backpropagation. The support set accuracy is greatly improved while the loss recovery accuracy meets the set threshold.

[0018] Step 8: Based on the above steps, establish the network model and data to train the neural network. After the training process is completed, deploy the system and build a compressed sensing reconstruction system based on the trained neural network.

[0019] In step 1, the generated sensing matrix is ​​the matrix product of the observation matrix and the sparse basis in the mathematical principle of compressed sensing. Its dimension is M×N, where M is the dimension of the sensing signal and N is the dimension of the original signal. The generation rule here satisfies the constrained equidistance condition required by the mathematical principle of compressed sensing.

[0020] In step 2, the generation of sparse signals conforms to the Bernoulli distribution or the normal distribution to simulate the real signal. The support set of the sparse signal is generally taken as the Bernoulli distribution or the normal distribution. The sparse signal used in the present invention conforms to the Bernoulli distribution, and the signal amplitude conforms to the normal distribution. Noise can be appropriately added in the simulation sensing process.

[0021] In step 3, the sparse signal that meets the requirements is directly matrix multiplied through the sensing matrix, which corresponds to the matrix multiplication of the observation matrix in the mathematical principle of compressed sensing and the signal that meets the sparse conditions in some transform domain.

[0022] In step 4, the EALISTA neural network model is constructed as where h θ (·) is the threshold function, θ is the trainable threshold, D is the sensing matrix, is the analysis matrix calculated in step 1, where y is the transposed matrix, y is the sensor signal, x is the recovered signal at each layer, and P is the trainable parameter matrix. The network is trained using supervised learning, and the network structure adopts a recurrent neural network-like structure. Each layer shares a portion of weights, while a small portion of weights is used independently by each layer. Each layer repeatedly inputs the sensor signal, which is used to calculate the output of the current layer together with the output of the previous layer. Leveraging the advantages of model-driven deep learning methods, the threshold function used in the model replaces the commonly used activation function, achieving both interpretability and high performance.

[0023] In step 5, the number of neurons in each layer of the network structure is closely related to the dimension of the sparse signal. Each layer of the neural network is regarded as an iteration in the iterative shrinkage threshold algorithm model. The result of each iteration is the recovery signal calculated in this iteration. Therefore, the number of neurons in each layer of the neural network is the dimension of the sparse signal to be restored. We use the output of the last layer of the neural network, that is, the recovery signal output by the last iteration as the output of the entire network, compare it with the label, and use their error as the loss function of the network for training.

[0024] In step 6, we use the mean squared error of the recovered signal and label described in step 5 as the loss function of the network, use the gradient descent optimizer, and perform backpropagation on the entire network to train the network weights.

[0025] In step 7, we replace the mean square error function in step 6 with our own custom error function. This error function increases the penalty for non-zero elements of the network's restored signal when the non-support set is non-zero, thereby suppressing the shortcoming of the mean square error only considering the error value of the overall signal without considering the support set position. The custom loss function expression is Where L1(·) and L0(·) are Where β is the penalty coefficient used to balance error and accuracy, K is the sum of the elements contained in the support set, and S is the support set of the non-zero elements of the true sparse signal.

[0026] In step 8, the training is offline training and the system deployment is online deployment. The offline training is the selection of network training data, network structure and network training parameters in steps 1-7, and the online deployment is to apply the trained network online.

[0027] An EALISTA compressed sensing reconstruction system driven by the ISTA model for the above method includes an analysis preprocessing unit and a neural network training unit. The analysis preprocessing unit needs to unify the sensing matrix at both ends of the signal transmission and reception. This sensing matrix is ​​an important component of compressed sensing reconstruction. The receiving end precalculates the analysis matrix based on the sensing matrix for subsequent sparse signal recovery; the neural network training unit mainly uses the improved neural network EALISTA to train network weights that can meet the needs of sparse signal recovery, and saves the trained neural network weight model for configuration in a specific system to complete the task of sparse signal recovery.

[0028] This solution includes multiple components, including a traditional iterative shrinkage threshold algorithm for compressed sensing reconstruction. This method pre-calculates an analysis matrix using a known sensor matrix, simplifies the training structure through the analysis matrix method, and improves the network structure based on this. This algorithm uses a quasi-recurrent neural network network structure, repeatedly inputting sensor signals at each layer. Based on the reconstruction model, the activation function is replaced by a threshold function, increasing interpretability while optimizing recovery performance, making the recovered sparse signal more accurate. Combined with a proposed custom loss function, the sparse signal support set can be further refined, facilitating the application of the reconstruction algorithm to systems sensitive to the position of the sparse signal support set. This method uses a multi-layer neural network, with the number of nodes in each layer matching the signal dimension to be recovered. Compared to conventional neural networks, the number of nodes is much smaller, enabling excellent recovery performance with a smaller-dimensional weight matrix. Each layer corresponds to each iteration of the model, effectively combining the excellent recovery performance of the traditional iterative shrinkage threshold algorithm and demonstrating the superior performance of model-driven neural networks.

[0029] As can be seen from the above description, this solution, based on the traditional compressed sensing reconstruction algorithm, applies deep learning, an artificial intelligence concept, to compressed sensing reconstruction. This approach optimizes the performance of traditional reconstruction algorithms through a model-driven architecture. Due to the limitations of traditional reconstruction algorithms, high numbers of algorithm iterations are required to achieve the required accuracy. However, this invention utilizes a deep learning network design, replacing each iteration with the network, making each iteration more precise and achieving better recovery performance with fewer iterations. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the structural diagram of the EALISTA compressed sensing reconstruction system driven by the designed model.

[0031] Figure 2 This is the neural network structure diagram of each layer of the designed system.

[0032] Figure 3 Comparison chart of loss function values ​​of different neural networks.

[0033] Figure 4 Comparison chart of normalized mean square error values ​​of different neural networks.

[0034] Figure 5 Comparison chart of normalized mean square error values ​​of different neural networks.

[0035] Figure 6 Comparison chart of the accuracy of different neural network support sets. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the specific embodiments of the present invention to clearly and completely describe the technical solutions in the specific embodiments of the present invention. Obviously, the specific embodiment described is only one specific embodiment of the present invention, not all specific embodiments. Based on the specific embodiments of the present invention, all other specific embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0037] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "comprising" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations;

[0038] It can be seen from the accompanying drawings that the EALISTA compressed sensing reconstruction method driven by the ISTA model of the present invention includes the following steps: step 1: according to the principle of compressed sensing, randomly generate a column-normalized Gaussian matrix as a sensing matrix, and calculate the corresponding analysis matrix according to the sensing matrix; step 2: based on the sensing matrix generated in step 1, randomly generate a sparse signal that meets the requirements, and generate a sensing signal by passing the sparse signal through the sensing matrix; step 3: based on the sparse signal generated in step 2 as a label, and the corresponding sensing signal generated in step 2 as training data, build a data set; step 4: build an EALISTA neural network based on the data set described in step 3, and the network structure is a quasi-cyclic neural network structure. Each layer of the network repeatedly inputs the training data generated in step 3, and each layer of the network corresponds to one iteration of the traditional iterative shrinkage threshold algorithm, so the output result of each layer of the network can be regarded as a reconstruction of the sparse signal, and the output of the upper network is passed into the lower network as iterative data, and each time The input is calculated together, and the activation function of each layer of the network is replaced by the threshold function of the iterative shrinkage threshold algorithm; Step 5: Based on the network described in Step 4, the error between the restored signal and the label output by the last layer of the neural network is used as the loss function, and the network weights are trained using the back propagation method; Step 6: Based on Step 5, the loss function adopts the mean square error of the restored signal and the label, the optimizer adopts the gradient descent optimizer, and the training is based on back propagation; Step 7: Based on Step 5, we proposed a new custom loss function to improve the accuracy of the restored signal support set. The loss function adopts the custom error function of the restored signal and the label, the optimizer adopts the gradient descent optimizer, and the training is based on back propagation, which can greatly improve the support set accuracy with a slight loss of recovery accuracy; Step 8: Based on the above steps, the network model and data are established to train the neural network. After the training process is completed, the system is deployed to build a compressed sensing reconstruction system based on the trained neural network.

[0039] In step 1, the generated sensing matrix is ​​the matrix product of the observation matrix and the sparse basis in the mathematical principle of compressed sensing. Its dimension is M×N, where M is the dimension of the sensing signal and N is the dimension of the original signal. The generation rule here satisfies the constrained equidistance condition required in the mathematical principle of compressed sensing. In step 2, the generation of sparse signals conforms to the Bernoulli distribution or normal distribution to simulate real signals. The support set of sparse signals generally takes the Bernoulli distribution or the normal distribution. The sparse signal used in the experiment conforms to the Bernoulli distribution, and the signal amplitude conforms to the normal distribution. Noise can be appropriately added in the simulation sensing process. In step 3, the sparse signal that meets the requirements is directly matrix multiplied by the sensing matrix, which corresponds to the matrix multiplication of the observation matrix in the mathematical principle of compressed sensing and the signal that meets the sparse conditions in certain transform domains. In step 4, the EALISTA neural network model is built as where h θ(·) is the threshold function, θ is the trainable threshold, D is the sensing matrix, is the analysis matrix calculated in step 1, Its transposed matrix, y is the sensor signal, x is the restored signal of each layer, and P is the trainable parameter matrix. The network training method adopts the supervised learning method, and the network structure adopts a quasi-cyclic neural network structure. Each layer of the network shares a part of the weights, and another small part of the weights is used independently by each layer of the network. Each layer of the network repeatedly inputs the sensor signal, which is used to calculate the output of the current layer of the network together with the output of the previous layer of the network. By taking advantage of the model-driven deep learning method, the threshold function used in the model replaces the commonly used activation function, which is both interpretable and has good performance. In step 5, the number of neurons in each layer of the network structure is closely related to the dimension of the sparse signal. Each layer of the neural network is regarded as an iteration in the iterative shrinkage threshold algorithm model. The result of each iteration is the restored signal calculated by this iteration. Therefore, the number of neurons in each layer of the neural network is the dimension of the sparse signal to be restored. We use the output of the last layer of the neural network, that is, the restored signal output by the last iteration as the output of the entire network, and compare it with the label. Their error is used as the loss function of the network for training. In step 6, we use the mean square error of the recovered signal and label described in step 5 as the loss function of the network, use the gradient descent optimizer, and perform backpropagation on the entire network to train the network weights. In step 7, we replace the mean square error function in step 6 with our own custom error function. This error function increases the penalty for non-zero elements of the network's recovered signal when the non-support set is non-zero, thereby suppressing the shortcoming of the mean square error only considering the error value of the overall signal without considering the position of the support set. The custom loss function expression is Where L1(·) and L0(·) are Where β is a penalty coefficient used to balance error and accuracy, K is the sum of the elements in the support set, and S is the support set of the nonzero elements of the true sparse signal. In step 8, the design is divided into offline training and online deployment. The offline training is similar to the selection of network training data, network structure, and network training parameters in steps 1-7, and the online deployment is the online application of the trained network.

[0040] The overall method in this scheme adopts a quasi-cyclic neural network structure, in which each layer of the neural network corresponds to an iteration in the traditional neural network recovery algorithm. The input of each layer of the neural network is fixed to the sensor signal y. Each layer of the neural network shares the weight matrix P but each layer of the neural network also has its own independent weight θ. The output of the last layer of the neural network is used as the output of the overall recovery algorithm. The input of each layer of the neural network needs to be preprocessed according to the steps defined by the model. The preprocessing process is as follows Figure 2 As shown in the figure, x k-1It is the output of the previous layer of network, and its dimension is the same as that of the original signal and the restored signal. The analysis matrix dimension is N×M, D is the sensing matrix generated in step 1, y is the sensing signal, and its dimension is M. The dimension of the signal after preprocessing is N. It serves as the input layer of the neural network, and is used as the output of this layer of neural network after being acted upon by the weight matrix P with dimension N×N and the threshold function.

[0041] Evaluation indicators of inventions such as Figure 3 、 4 , 5 and 6. As described Figure 3 As shown, Figure 3 The vertical axis of the coordinate axis is the mean square error loss value of the neural network, and the horizontal axis is the signal-to-noise ratio of the sensing process. As shown in the figure, the loss value of the invention is lower than that of the existing network, and it can still maintain better recovery accuracy than the existing network under low signal-to-noise ratio conditions, as described Figure 4 As shown, Figure 4 Coordinate axis and Figure 3 Consistent, Figure 4 The comparison is made with the loss value under the proposed custom loss function. As shown in the figure, due to the defects of the LISTA network itself, the custom loss function has a better optimization effect on the LISTA network. Under low signal-to-noise ratio conditions, the performance of the LISTA network is similar to or even better than the proposed EALISTA network. The main function of the custom loss function is to optimize the support set accuracy, as described above. Figure 5 As shown, Figure 5 The difference in loss values ​​under the mean square error loss function and the custom loss function is compared. The ones with the "mse" suffix are under the mean square error, and the ones without the suffix are under the custom loss function. The vertical axis of the coordinate axis is the decibel form of the normalized mean square error loss value of the neural network, and the horizontal axis is the signal-to-noise ratio of the sensing process. As shown in the figure, under the mean square error, the invention can reduce the normalized mean square error value by 13dB at a signal-to-noise ratio of 40dB. Due to the "penalty" added in the custom loss function, the errors of ALISTA and EALISTA are slightly increased, while the error of LISTA, as mentioned earlier, is reduced. Figure 6 As shown, the vertical axis is the accuracy of the neural network restored signal, and the horizontal axis is the signal-to-noise ratio of the sensing process. As shown in the figure, the accuracy of the restored signal support set of this invention is high, and the custom loss function proposed by this invention can greatly improve the accuracy of each neural network restored signal support set.

[0042] Although specific embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these specific embodiments without departing from the principles and spirit of the invention, and the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An ISTA model-driven EALISTA compressed sensing reconstruction method, characterized by The following steps are included: Step 1: Based on the principle of compressed sensing, a column-normalized Gaussian matrix is ​​randomly generated as the sensing matrix, and the corresponding analysis matrix is ​​calculated based on the sensing matrix; Step 2: Based on the sensing matrix generated in step 1, randomly generate a sparse signal that meets the requirements, and pass the sparse signal through the sensing matrix to generate a sensing signal; Step 3: Use the sparse signal generated in step 2 as the label and the corresponding sensor signal generated in step 2 as the training data to build a data set; Step 4: Build an EALISTA neural network based on the data set described in step 3. The network structure is a quasi-cyclic neural network structure. Each layer of the network repeatedly inputs the training data generated in step 3, and each layer of the network corresponds to an iteration of the traditional iterative shrinkage threshold algorithm. Therefore, the output result of each layer of the network is a reconstruction of the sparse signal. The output of the upper layer network is passed to the lower layer network as iterative data. In the lower layer network, the output of the upper layer network is calculated together with the input training data each time. The activation function of each layer of the network is replaced by the threshold function of the iterative shrinkage threshold algorithm. Step 5: Based on the network described in step 4, the error between the restored signal output by the last layer of neural network and the label is used as the loss function, and the network weights are trained using the back propagation method; Step 6: Based on step 5, the loss function uses the mean square error between the restored signal and the label, and the optimizer uses the gradient descent optimizer, which is trained based on backpropagation. Step 7: Based on step 5, a new loss function is used to improve the accuracy of the restored signal support set. The loss function uses a custom error function of the restored signal and the label. The optimizer uses a gradient descent optimizer and is trained based on backpropagation to improve the support set accuracy while ensuring that the loss recovery accuracy meets the set threshold. Step 8: Based on the above steps, establish the network model and data to train the neural network. After the training process is completed, deploy the system and build a compressed sensing reconstruction system based on the trained neural network.

2. The EALISTA compressed sensing reconstruction method driven by the ISTA model according to claim 1, characterized in that , In step 1, the generated sensing matrix is ​​the matrix product of the observation matrix and the sparse basis in the mathematical principle of compressed sensing. Its dimension is M×N, where M is the dimension of the sensing signal and N is the dimension of the original signal. The generation rule here satisfies the constrained equidistance condition required by the mathematical principle of compressed sensing.

3. The EALISTA compressed sensing reconstruction method driven by the ISTA model according to claim 1, characterized in that , In step 2, the generation of sparse signals conforms to the Bernoulli distribution or the normal distribution to simulate the real signal. The support set of the sparse signal takes the Bernoulli distribution or the normal distribution. The sparse signal used in the experiment conforms to the Bernoulli distribution, and the signal amplitude conforms to the normal distribution.

4. The EALISTA compressed sensing reconstruction method driven by the ISTA model according to claim 1, characterized in that , In step 3, the sparse signal that meets the requirements is matrix multiplied through the sensing matrix, which corresponds to the matrix multiplication of the observation matrix in the mathematical principle of compressed sensing and the signal that meets the sparse conditions in some transform domain.

5. The EALISTA compressed sensing reconstruction method driven by the ISTA model according to claim 1, characterized in that , In step 4, the EALISTA neural network model is constructed as where h θ (·) is the threshold function, θ is the trainable threshold, D is the sensing matrix, is the analysis matrix calculated in step 1, is its transposed matrix, y is the sensor signal, x is the restored signal of each layer, P is the trainable parameter matrix, the network training method adopts supervised learning, the network structure adopts a quasi-recurrent neural network structure, each layer of the network shares a part of the weights, and the other part of the weights is used independently by each layer of the network. Each layer of the network repeatedly inputs the sensor signal, which is used to calculate the output of the current layer of the network together with the output of the previous layer of the network. Taking advantage of the model-driven deep learning method, the threshold function used in the model replaces the commonly used activation function, which is both interpretable and has good performance.

6. The EALISTA compressed sensing reconstruction method driven by the ISTA model according to claim 1, characterized in that , In step 5, the number of neurons in each layer of the network structure is closely related to the dimension of the sparse signal. Each layer of the neural network is regarded as an iteration in the iterative shrinkage threshold algorithm model. The result of each iteration is the recovery signal calculated in this iteration. Therefore, the number of neurons in each layer of the neural network is the dimension of the sparse signal to be restored. The output of the last layer of the neural network, that is, the recovery signal output by the last iteration is used as the output of the entire network, which is compared with the label, and the error after comparison is used as the loss function of the network for training.

7. The ISTA model-driven EALISTA compressed sensing reconstruction method according to claim 1, characterized in that , In step 6, we use the mean squared error of the recovered signal and label described in step 5 as the loss function of the network, use the gradient descent optimizer, and perform backpropagation on the entire network to train the network weights.

8. The ISTA model-driven EALISTA compressed sensing reconstruction method according to claim 1, characterized in that , In step 7, we replace the mean square error function in step 6 with a custom error function. This custom error function increases the penalty for non-zero elements of the network's restored signal non-support set, thereby suppressing the shortcoming that the mean square error only considers the error value of the overall signal without considering the support set position. The custom loss function expression is Where L1(·) and L0(·) are Where β is the penalty coefficient used to balance error and accuracy, K is the sum of the elements contained in the support set, and S is the support set of the non-zero elements of the true sparse signal.

9. The ISTA model-driven EALISTA compressed sensing reconstruction method according to claim 1, characterized in that , In step 8, the training is offline training and the system deployment is online deployment. The offline training is the selection of network training data, network structure and network training parameters in steps 1-7, and the online deployment is to apply the trained network online.

10. An ISTA model-driven EALISTA compressed sensing reconstruction system according to any one of claims 1 to 9, characterized in that , Including analysis preprocessing unit and neural network training unit, The analysis preprocessing unit needs to unify the sensor matrix at both ends of the signal transmission and reception. This sensor matrix is ​​a component of compressed sensing reconstruction. The receiving end pre-calculates the analysis matrix based on the sensor matrix for subsequent sparse signal recovery. The neural network training unit uses the improved neural network EALISTA to train the network weights that can meet the needs of sparse signal recovery, and saves the trained neural network weight model for configuration in a specific system to complete the task of sparse signal recovery.

Images