A lightweight image reconstruction method based on an AMP network

Abstract

The application discloses a kind of light-weight image reconstruction method, system and computer readable storage medium based on AMP network, method includes the following steps: S1: obtaining two-dimensional image, two-dimensional image is segmented into the image sub-block of pre-set number, size fixed by block function, image sub-block is converted into one-dimensional vector signal;S2: construct light-weight AMP network, the light-weight AMP network includes the sampling framework, initialization module and multilayer reconstruction block connected in turn;S3: one-dimensional vector signal is input to light-weight AMP network and output reconstructed image.This application constructs light-weight AMP network, introduces full map convolution in multilayer reconstruction block to make full use of image global information, to remove artifact, so as to not need to introduce additional artifact removal module, so that under the premise of reconstruction quality is equivalent even more optimal, reconstruction time and model parameter quantity are greatly reduced, reduce the resources and time required for image reconstruction.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and more specifically, to a lightweight image reconstruction method and system based on AMP networks, and a computer-readable storage medium. Background Technology

[0002] With the rapid development of the Internet, large-capacity unstructured data such as high-definition images, videos, and GIFs have become widespread. Coupled with people's continuous pursuit of image quality, the transmission process places extremely high demands on channel bandwidth, processing speed, and memory resource consumption. Compressive image sensing (CIS) is a technology that can sample and reconstruct image signals at ultra-low sampling rates. Thanks to its simple encoding mechanism, CIS is usually used for image information acquisition and data compression in resource-constrained scenarios. However, for the decoding end, with the rapid development of mobile applications, more and more resource-constrained scenarios require the resources and time required for image reconstruction to be minimized. Therefore, we need to reduce the complexity of the reconstruction algorithm, model size, and reconstruction time as much as possible without sacrificing reconstruction quality.

[0003] Compressive sensing (CS) is a signal processing technique that uses ultra-low sampling rates to compress and reconstruct signals. In 2006, Candès, Romberg, Tao, and Donoho et al. pointed out that when a signal has a sparse representation, it can be accurately reconstructed from a series of linear, non-adaptive observations. For a K-stage sparse signal x∈R... N At the encoding end, we can use a linear mapping y = Φx to simultaneously compress and sample the signal, obtaining the observed value y ∈ R. N , where Φ∈R M×N (M << N) is the sampling matrix. Then we can transmit the extremely short observation value y to the decoding end. After the decoding end obtains y, it can accurately reconstruct x using the corresponding algorithm.

[0004] Although natural image signals are not sparse, we can perform wavelet transforms on them to obtain approximate coefficient representations. Wavelet transform converts the image signal into high-frequency and low-frequency components. The low-frequency component provides a coarse-scale approximation of the image, while the high-frequency component fills in the image details. When calculating the wavelet coefficients of a natural image, most coefficients are very small; therefore, we can set the small coefficients to zero to obtain an approximate representation of the image. Thus, for a natural image signal s∈R... N Through the sampling matrix Φ∈R M×N Compressed sampling y = Φs is performed to obtain the measured value y ∈ R. MAt this point, we can use y to reconstruct an approximate sparse representation of the image x∈R. N Through inverse wavelet transform The image signal is reconstructed, where Ψ∈R N×N is the inverse wavelet transform matrix.

[0005] Existing solutions can be mainly divided into two categories. The first category is non-iterative neural networks. This type of solution first builds an initial neural network to map y to x, and then utilizes the powerful learning capabilities of deep learning (DL) to train a high-quality neural network and a better sampling matrix using a large amount of data. This type of solution has fast reconstruction speed and high image reconstruction quality. However, because this type of solution treats the neural network as a "black box" during training, its mathematical interpretability is weak, and it requires sufficient data for training. The second category is deep unfolding methods. This type of solution unfolds the traditional compressed sensing iterative algorithm into the form of a neural network. It not only has the advantage of strong mathematical interpretability of traditional iterative algorithms, but also inherits the excellent learning capabilities of neural networks. It uses training data to obtain a signal reconstruction model with fast speed and high reconstruction quality.

[0006] Because compressing and sampling an entire image results in a very large number of sampling matrix parameters, sampling computation, and reconstruction computation, this type of scheme first needs to divide the image into several non-overlapping image sub-blocks to reduce computational cost. Then, each sub-block is sampled separately. However, the reconstruction of each image sub-block is carried out independently, which generates a large number of blocking artifacts. Therefore, it is often necessary to introduce an additional artifact removal module after block-by-block reconstruction. This greatly increases the computational complexity and number of parameters of the model, causing the model's storage space and reconstruction time to rise sharply. This is not conducive to the rapid training of the model or image reconstruction in resource-constrained environments.

[0007] Existing technology discloses an image reconstruction method and system. The image reconstruction method includes: step a: performing spectral separation on the image to obtain the spectral components of the image; step b: performing peak point search on the spectral components to obtain the direction and spatial frequency of the illumination light; step c: calculating the modulation and initial phase of the illumination light using a linear fitting algorithm, correcting the spectral components using the modulation and initial phase of the illumination light, and normalizing the coefficients of the spectral components; step d: performing deconvolution and translation processing on the normalized spectral components, and combining the deconvolutioned and translated spectral components to obtain the reconstructed image result. Although this scheme involves image reconstruction, it utilizes spectral components and does not achieve lightweight image reconstruction. Summary of the Invention

[0008] This invention provides a lightweight image reconstruction method and system based on AMP networks, as well as a computer-readable storage medium. While ensuring the quality of image reconstruction, it reduces reconstruction time and the number of model parameters, minimizing the resources and time required for reconstruction.

[0009] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0010] The first aspect of this invention provides a lightweight image reconstruction method based on AMP networks, comprising the following steps:

[0011] S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals;

[0012] S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence;

[0013] S3: Input a one-dimensional vector signal into a lightweight AMP network to output the reconstructed image.

[0014] Furthermore, the specific process of step S1 is as follows:

[0015] Obtain a 2D image and divide it into L non-overlapping image sub-blocks {X1,…,X2,…} using a block function. L},in Use vectorization functions to convert image sub-patterns into one-dimensional vector signals {x1,…,x}. l ,…,X L},in Let {x1,…,x l ,…,x L} Integrate into matrix form, denoted as

[0016] Furthermore, in the sampling framework, a sampling matrix is ​​used to linearly map the one-dimensional vector signal x to complete compressed sampling, where the expression for the linear mapping is y = Ax, and the sampling process is represented as follows:

[0017] y=ASV(X) (1)

[0018] SV() represents the vectorization function, and A represents the sampling matrix. y represents the sampled value, y∈R M×L .

[0019] Furthermore, the initialization module receives the sampled values ​​from the sampling frame and initializes them to obtain the initial image signal, as shown in the following expression:

[0020] x 0 =A T y (2)

[0021] Wherein, y is the sampled value, and the initial image signal is... This is the transpose of the sampling matrix A.

[0022] Furthermore, the multi-level reconstruction block processing procedure is as follows:

[0023] The initialization image signal output from the initialization module is input into the multi-layer reconstruction block to reconstruct the image iteratively. The iterative process of the k-th layer of the l-th image sub-block is as follows:

[0024]

[0025]

[0026] Among them, y l ∈R M and Represent the measurement value of the l-th image sub-block and the output of the (k-1)-th layer, respectively, η k (·) and α k These are the nonlinear iterative function and step size of the k-th layer, respectively;

[0027] y l =Ax l and Substitute into In the middle, there are:

[0028]

[0029] In equation (5) above, x represents the original signal x. l The lossless expression, where only a sufficiently good fit is needed. This will allow us to obtain a reconstructed signal that is sufficiently similar to the original signal x;

[0030] use To fit in For iteration value The two-dimensional image signal form, F k (·) is a four-layer convolutional network. The kernel size and padding size of each layer are 3×3 and 1×1, respectively. The first three convolutional layers have bias and ReLU (Rectified Linear Unit) layers, while the last convolutional layer does not have bias and ReLU layers.

[0031] In summary, the process of the k-th layer iteration can be summarized as follows:

[0032]

[0033]

[0034] Stitch together all the image sub-blocks from the output of the k-th layer to form the complete image X. k-1 ∈R H×W As F k Given the input (·), the k-th layer iteration is represented as:

[0035] v k-1 =y-Ax k-1 (8)

[0036]

[0037]

[0038] in For L image sub-blocks A matrix formed by integration.

[0039] To further reduce the amount of matrix operations in the iteration, substitute (8) into (10) and derive as follows:

[0040]

[0041]

[0042] in

[0043] Furthermore, the multi-layer reconstruction block parameters include: sampling matrix A, step size Θ of each layer. α ={α1,α2,……,α K} and the parameters Θ of the K convolutional networks F ={F1,F2,……,F K};

[0044] Using mean squared error as the loss function for multi-layer reconstruction blocks allows the network to better reduce the original signal X and the reconstructed signal after training. Error between:

[0045]

[0046] Where, N b X is the number of images in the training set. i Let be the i-th training image.

[0047] A second aspect of the present invention provides a lightweight image reconstruction system based on an AMP network. The system includes a memory and a processor. The memory includes a lightweight image reconstruction method program based on an AMP network. When executed by the processor, the lightweight image reconstruction method program based on an AMP network performs the following steps:

[0048] S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals;

[0049] S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence;

[0050] S3: Input a one-dimensional vector signal into a lightweight AMP network to output the reconstructed image.

[0051] Furthermore, the specific process of step S1 is as follows:

[0052] Obtain a 2D image and divide it into L non-overlapping image sub-blocks {X1,…,X2,…} using a block function. L},in Use vectorization functions to convert image sub-patterns into one-dimensional vector signals {x1,…,x}. l ,…,x L},in Let {x1,…,x l ,…,x L} can be integrated into a matrix form, denoted as

[0053] Furthermore, in the sampling framework, a sampling matrix is ​​used to linearly map the one-dimensional vector signal x to complete compressed sampling, where the expression for the linear mapping is y = Ax, and the sampling process is represented as follows:

[0054] y=ASV(X) (1)

[0055] SV() represents the vectorization function, and A represents the sampling matrix. y represents the sampled value, y∈R M×L .

[0056] A third aspect of the present invention provides a computer-readable storage medium including a lightweight image reconstruction method program based on an AMP network. When the lightweight image reconstruction method program based on an AMP network is executed by a processor, it implements the steps of the aforementioned lightweight image reconstruction method based on an AMP network.

[0057] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0058] This invention constructs a lightweight AMP network and introduces full-image convolution in multi-layer reconstruction blocks to fully utilize global image information, thereby removing artifacts. This eliminates the need for an additional artifact removal module, significantly reducing reconstruction time and model parameter count while maintaining comparable or even better reconstruction quality, thus minimizing the resources and time required for reconstruction. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the signal sampling process based on compressed sensing in the prior art of this invention.

[0060] Figure 2 This is a flowchart of a lightweight image reconstruction method based on AMP network according to an embodiment of the present invention.

[0061] Figure 3 This is a diagram of a lightweight AMP network architecture according to an embodiment of the present invention.

[0062] Figure 4 This is a network structure diagram of the k-th reconstruction layer in an embodiment of the present invention.

[0063] Figure 5 This is a comparison of reconstructed images at a sampling rate of 30% according to an embodiment of the present invention.

[0064] Figure 6 This is a comparison chart of model training time at a 10% sampling rate in an embodiment of the present invention. Detailed Implementation

[0065] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0066] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0067] Example 1

[0068] like Figure 2 As shown, the first aspect of the present invention provides a lightweight image reconstruction method based on AMP networks, comprising the following steps:

[0069] S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals;

[0070] It should be noted that in this application, a two-dimensional image is first obtained, denoted as X∈R. H×W The two-dimensional image is divided into L non-overlapping image sub-blocks {X1,…,X2,…} of size B×B using the block function S(·). L},in The image sub-blocks are converted into one-dimensional vector signals {x1,…,x} using the vectorization function V(·). l ,…,x L},in Let {x1,…,x l ,…,x L} Integrate into matrix form, denoted as The above process can be integrated into the splitting and vectorization function SV(·), so the whole process can be summarized as x = SV(X).

[0071] Correspondingly, the process of reassembling L one-dimensional signals x into a complete two-dimensional image signal X is represented by the image blocking and sub-block stitching function BM(·), which can be summarized as X = BM(x).

[0072] Example 2

[0073] This embodiment describes the lightweight AMP network in detail as follows:

[0074] S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence;

[0075] Figure 3 A lightweight AMP network architecture is shown.

[0076] It should be noted that the lightweight AMP network constructed in this invention includes a sampling framework and a reconstruction framework. The reconstruction framework includes an initialization module and multiple reconstruction blocks. In the sampling framework, a sampling matrix is ​​used to linearly map the one-dimensional vector signal x to complete compressed sampling, where the expression for the linear mapping is y = Ax.

[0077] A represents the sampling matrix. y represents the sampled value, y∈R M×L .

[0078] The initialization module receives the sampled values ​​from the sampling frame and initializes them to obtain the initial image signal, as shown in the following expression:

[0079] x 0 =A T y (2)

[0080] Wherein, y is the sampled value, and the initial image signal is... This is the transpose of the sampling matrix A.

[0081] The multi-level reconstruction block processing procedure is as follows:

[0082] The initialization image signal output from the initialization module is input into the multi-layer reconstruction block to reconstruct the image iteratively. The iterative process of the k-th layer of the l-th image sub-block is as follows:

[0083]

[0084]

[0085] Among them, y l ∈R M and Represent the measurement value of the l-th image sub-block and the output of the (k-1)-th layer, respectively, η k (·) and α k These are the nonlinear iterative function and step size of the k-th layer, respectively;

[0086] y l =Ax l and Substitute into In the middle, there are:

[0087]

[0088]

[0089] In equation (5) above, x represents the original signal x. l The lossless expression, where only a sufficiently good fit is needed. This will allow us to obtain a reconstructed signal that is sufficiently similar to the original signal x; when each element in A follows a Gaussian independent and identically distributed N(0,μ), It will follow a Gaussian distribution with variance of . Furthermore, the variance decreases with iteration.

[0090] In this application, the use of To fit in For iteration value The two-dimensional image signal form, F k (·) is a four-layer convolutional network. The kernel size and padding size of each layer are 3×3 and 1×1, respectively. The first three convolutional layers have bias and ReLU (Rectified Linear Unit) layers, while the last convolutional layer does not have bias and ReLU layers.

[0091] It should be noted that fitting is done using a convolutional network. The biggest advantage is that it makes full use of the excellent learning ability of deep learning. By extracting useful information from the training data, it learns the prior knowledge of the image, thereby reducing the design cost of the model for signal priors. The quality of image reconstruction actually depends on the training of the convolutional network. Therefore, as long as the model is trained properly, it can theoretically cope with image signals with various signal priors.

[0092] In summary, the process of the k-th layer iteration can be summarized as follows:

[0093]

[0094]

[0095] It should be noted that since reconstructing each image patch individually will produce artifacts, AMP-Net needs to add a convolutional network with the whole image as input after each patch reconstruction to remove artifacts. Taking a four-layer AMP-Net as an example, the image is processed alternately by four patch reconstruction modules and four artifact removal modules in multiple reconstruction layers. All eight modules need to perform convolution operations. Such stacked convolution operations will limit the network's running speed and greatly increase the number of network parameters and training cost.

[0096] Stitch together all the image sub-blocks from the output of the k-th layer to form the complete image X. k-1 ∈R H×W As F k With the input of (·), the information between image sub-blocks can be fully utilized to eliminate artifacts in the process of layer-by-layer reconstruction (7), so that no additional artifact removal is needed after block-by-block reconstruction. This means that only four modules are needed in the four-layer network to reconstruct an image without artifacts, which can greatly reduce reconstruction time, model parameter quantity and training cost.

[0097] Stitch together all the image sub-blocks from the output of the k-th layer to form the complete image X. k-1 ∈R H×W As F kGiven the input of (·) (i.e., after modifying the convolutional input to a full-image convolution during the block-by-block reconstruction process), the k-th layer iteration is represented as:

[0098] v k-1 =y-Ax k-1 (8)

[0099]

[0100]

[0101] in For L image sub-blocks A matrix formed by integration.

[0102] To further reduce the amount of matrix operations in the iteration, substitute (8) into (10) and derive as follows:

[0103]

[0104] in

[0105] Figure 4 The network structure of the k-th reconstruction layer is shown.

[0106] The parameters of the multi-layer reconstruction block include: sampling matrix A, step size Θ of each layer. α ={α1,α2,……,α K} and the parameters Θ of the K convolutional networks F ={F1,F2,……,F K};

[0107] Using mean squared error as the loss function for multi-layer reconstruction blocks allows the network to better reduce the original signal X and the reconstructed signal after training. Error between:

[0108]

[0109] Where, N b X is the number of images in the training set. i Let be the i-th training image.

[0110] S3: Input a one-dimensional vector signal into a lightweight AMP network to output the reconstructed image.

[0111] A second aspect of the present invention provides a lightweight image reconstruction system based on an AMP network. The system includes a memory and a processor. The memory includes a lightweight image reconstruction method program based on an AMP network. When executed by the processor, the lightweight image reconstruction method program based on an AMP network performs the following steps:

[0112] S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals;

[0113] S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence;

[0114] S3: Input a one-dimensional vector signal into a lightweight AMP network to output the reconstructed image.

[0115] Furthermore, the specific process of step S1 is as follows:

[0116] Obtain a 2D image and divide it into L non-overlapping image sub-blocks {X1,…,X2,…} using a block function. L},in Use vectorization functions to convert image sub-patterns into one-dimensional vector signals {x1,…,x}. l ,…,x L},in Let {x1,…,x l ,…,x L} can be integrated into a matrix form, denoted as

[0117] Furthermore, in the sampling framework, a sampling matrix is ​​used to linearly map the one-dimensional vector signal x to achieve compressed sampling, where the expression for the linear mapping is y = Ax, and A represents the sampling matrix. y represents the sampled value, y∈R M×L .

[0118] A third aspect of the present invention provides a computer-readable storage medium including a lightweight image reconstruction method program based on an AMP network. When the lightweight image reconstruction method program based on an AMP network is executed by a processor, it implements the steps of the aforementioned lightweight image reconstruction method based on an AMP network.

[0119] Example 3

[0120] This embodiment uses specific experiments to verify and analyze the results. In this embodiment, the reconstruction quality, reconstruction speed and model parameter quantity are compared with ISTA-Net+, CS-Net+ and 4-layer and 6-layer AMP-Net (denoted as AMP-Net-4 and AMP-Net-6, respectively).

[0121] Experimental setup

[0122] This embodiment primarily uses the BSDS500 and Set11 datasets. Each image in the BSDS500 training set (containing 200 images) is randomly split into 448 99×99 images to generate the experimental training set. The BSDS500 validation set (containing 200 images) is used for training validation, and the BSDS500 test set (containing 100 images) and Set11 (containing 11 images) are used for testing. Training uses Peak Signal-to-Noise Ratio (PSNR) and SSIM.

[0123] For the initialization settings of the model parameters, the image sub-block size B = 33, the number of layers of the lightweight AMP network (LightweightAMP-Net, LwAMP-Net) is set to 4 and 6 layers respectively (abbreviated as LwAMP-4 and LwAMP-6), the sampling matrix A is initialized with a random Gaussian matrix, and the parameters Θ of the convolutional network are... F Using random initialization, the step size Θ for each layer α Initialize to 1.

[0124] During training, the Adam optimizer was used for 100 epochs, with a batch size of 64 and a learning rate of 2×10⁻⁶. -4 All experiments were implemented in PyTorch, with the experimental platform configured as an Intel Core i7-9700 CPU and an RTX 3070 GPU.

[0125] Reconstruction quality comparison

[0126] In Table 1 and Figure 5 The results show that LwAMP's reconstruction quality is comparable to or even better than other methods. For example, at a sampling rate of 10%, LwAMP-6 has a PSNR 0.23 dB higher than AMP-6. Table 1 compares the PSNR and SSIM of the reconstructed images on the Set11 and BSDS500 test sets, respectively, with the second and third results marked in bold, underline, and italics, respectively.

[0127] Table 1 compares the PSNR and SSIM of reconstructed images on the Set11 and BSDS500 test sets.

[0128]

[0129] Comparison of reconstruction speed and model parameter count

[0130] As shown in Table 2, LwAMP-4 is 1.68-2.75 times faster than other methods in terms of reconstruction speed, and its parameter count is only half that of AMP-Net-4 with the same number of layers. Therefore, LwAMP can significantly reduce reconstruction time and model parameter count while ensuring reconstruction quality, so that the resources and time required for reconstruction are minimized.

[0131] Table 2 compares the number of model parameters and the CPU and GPU time for reconstructing a 256×256 image at a 50% sampling rate.

[0132]

[0133] Model training time comparison

[0134] like Figure 6 As shown, the training time of LwAMP-4 and LwAMP6 is only half that of AMP-Net4 and AMP-Net6, respectively. Compared with ISTA-Net+ and CSNet+, it also has a significant training speed advantage. This is due to the merging of the two modules in each iteration of the original AMP-Net into one module, which greatly reduces the amount of gradient calculation during training.

[0135] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A lightweight image reconstruction method based on AMP networks, characterized in that, Includes the following steps: S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals; S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence; S3: Input a one-dimensional vector signal into the lightweight AMP network and output the reconstructed image; The multi-level reconstruction block processing procedure is as follows: The initialization image signal output from the initialization module is input to the multi-layer reconstruction block to reconstruct the image iteratively. The first image sub-block The iterative process of the layer is as follows: (3) (4) in, and They represent the first The measurement value of the first image sub-block and the first The output of the layer, and The first The nonlinear iterative function and step size of the layer; Will and Substitute into In the middle, there are: (5) In equation (5) above, the original signal is... The lossless expression, where only a sufficiently good fit is needed. This will allow you to obtain a signal similar to the original signal. Sufficiently similar reconstructed signals; use To fit ,in For iteration value Two-dimensional image signal form, This is a four-layer convolutional network, with the kernel size and padding size of each layer being as follows: and The first three convolutional layers have bias and ReLU layers, while the last convolutional layer does not have bias and ReLU layers. In summary, the first The process of layer iteration can be summarized as follows: (6) (7) The first All image sub-blocks output by the layer are stitched together to form a complete image. As The input, then the first Layer iteration is represented as: (8) (9) (10) in for Image sub-blocks An integrated matrix; To further reduce the amount of matrix operations in the iteration, substitute (8) into (10) and derive as follows: (11) in ; Multi-layer reconstruction block parameters include: sampling matrix Step size of each layer and The parameters of a convolutional network ; Using mean squared error as the loss function for multi-layer reconstruction blocks allows the network to better reduce the original image signal after training. and reconstructed image signals Error between: (12) in, The number of images in the training set. For the first Zhang training image signal.

2. The lightweight image reconstruction method based on AMP network according to claim 1, characterized in that, The specific process of step S1 is as follows: Obtain a 2D image and segment it into L non-overlapping segments of size L using a block function. Image sub-blocks { ,in The image sub-patterns are converted into one-dimensional vector signals using a vectorization function. ,in ,Will Integrate into matrix form, denoted as .

3. The lightweight image reconstruction method based on AMP network according to claim 1, characterized in that, In the sampling framework, a sampling matrix is ​​used to represent a one-dimensional vector signal. Compressed sampling is achieved by performing a linear mapping, where the expression for the linear mapping is: The sampling process is represented as follows: (1) Represents the sampling matrix , Indicates the sampled value. .

4. The lightweight image reconstruction method based on AMP network according to claim 1, characterized in that, The initialization module receives the sampled values ​​from the sampling frame and initializes them to obtain the initial image signal, as shown in the following expression: （2） in, Sampled values, initial image signal , Sampling matrix The transpose of .

5. A lightweight image reconstruction system based on AMP networks, characterized in that, The system includes: a memory and a processor. The memory includes a lightweight image reconstruction method program based on AMP networks. When the processor executes the lightweight image reconstruction method program based on AMP networks, it performs the following steps: S1: Acquire a two-dimensional image, divide the two-dimensional image into a preset number of image sub-blocks of fixed size using a block function, and convert the image sub-blocks into one-dimensional vector signals; S2: Construct a lightweight AMP network, which includes a sampling framework, an initialization module, and a multi-layer reconstruction block connected in sequence; S3: Input a one-dimensional vector signal into the lightweight AMP network and output the reconstructed image; The multi-level reconstruction block processing procedure is as follows: The initialization image signal output from the initialization module is input to the multi-layer reconstruction block to reconstruct the image iteratively. The first image sub-block The iterative process of the layer is as follows: (3) (4) in, and They represent the first The measurement value of the first image sub-block and the first The output of the layer, and The first The nonlinear iterative function and step size of the layer; Will and Substitute into In the middle, there are: (5) In equation (5) above, the original signal is... The lossless expression, where only a sufficiently good fit is needed. This will allow you to obtain a signal similar to the original signal. Sufficiently similar reconstructed signals; use To fit ,in For iteration value Two-dimensional image signal form, This is a four-layer convolutional network, with the kernel size and padding size of each layer being as follows: and The first three convolutional layers have bias and ReLU layers, while the last convolutional layer does not have bias and ReLU layers. In summary, the first The process of layer iteration can be summarized as follows: (6) (7) The first All image sub-blocks output by the layer are stitched together to form a complete image. As The input, then the first Layer iteration is represented as: (8) (9) （10）(10) in for Image sub-blocks An integrated matrix; To further reduce the amount of matrix operations in the iteration, substitute (8) into (10) and derive as follows: (11) in ; Multi-layer reconstruction block parameters include: sampling matrix Step size of each layer and The parameters of a convolutional network ; Using mean squared error as the loss function for multi-layer reconstruction blocks allows the network to better reduce the original image signal after training. and reconstructed image signals Error between: (12) in, The number of images in the training set. For the first Zhang training image signal.

6. A lightweight image reconstruction system based on an AMP network according to claim 5, characterized in that, The specific process of step S1 is as follows: Obtain a 2D image and segment it into L non-overlapping segments of size L using a block function. Image sub-blocks ,in The image sub-patterns are converted into one-dimensional vector signals using a vectorization function. ,in ,Will Integrate into matrix form, denoted as .

7. A lightweight image reconstruction system based on an AMP network according to claim 5, characterized in that, In the sampling framework, a sampling matrix is ​​used to represent a one-dimensional vector signal. Compressed sampling is achieved by performing a linear mapping, where the expression for the linear mapping is: , Represents the sampling matrix , Indicates the sampled value. .

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a lightweight image reconstruction method program based on AMP networks. When the lightweight image reconstruction method program based on AMP networks is executed by a processor, it implements the steps of a lightweight image reconstruction method based on AMP networks as described in any one of claims 1 to 4.

Images