A compression and reconstruction method for industrial images

Abstract

The present invention discloses a compression and reconstruction method for industrial images. The steps include: graying the industrial image, cutting it into an image to be processed, extracting sparse features through wavelet transform, and obtaining a sparse image; establishing a restoration model for the sparse image, iteratively processing and outputting the final restoration coefficient of the sparse image; processing the restoration coefficient through inverse wavelet transform to obtain a restored image; splicing each restored image to obtain a spliced ​​image, that is, obtaining a reconstruction result of the industrial image. The method of the present invention can effectively compress large-scale images in industrial production, improve transmission efficiency, and can quickly and accurately restore the compressed image in the reconstruction after compression for later detection.

Description

Technical Field

[0001] The invention relates to a compression reconstruction method, in particular to a compression reconstruction method for industrial images. Background Art

[0002] With the rapid development of industrial digitalization and intelligence, we have entered the era of "big data". With it comes the dependence of all walks of life on big data. Therefore, improving data processing capabilities has increasingly become the main way for all walks of life to promote the creation of industrial value, improve quality, reduce costs and increase benefits. Taking industrial production as an example, with the introduction of the concept of intelligent manufacturing, it has become an inevitable trend to replace manual visual inspection with machine vision inspection to achieve intelligent and automated production. However, due to the improvement of industrial camera pixels, the amount of data that needs to be processed in modern machine vision inspection has also increased, including high-dimensional data processing. If you want to solve this problem, the most direct and effective way is to improve the computing power hardware, but this means that the production cost will increase significantly. In order to solve this bottleneck, some experts have proposed in recent years to replace local computing power equipment with cloud server computing power to supplement computing power, thereby solving the contradiction between improving computing power and reducing production costs. But at the same time, it will face another problem, that is, how to transmit the detection data more effectively, that is, how to solve the high cost, high redundancy, low efficiency and other problems faced by large-scale image transmission. Summary of the invention

[0003] As a sparse reconstruction method based on Bayesian theorem, the Sparse Bayesian Learning (SBL) algorithm model has obvious advantages over other sparse representation models in reducing the risk of overfitting of the reconstructed prediction model, accelerating the algorithm's solution speed, and improving the robustness of the algorithm, relying on its statistical probability and Bayesian theoretical foundation, and provides a possibility for solving the above problems. In addition, Generalized Approximate Message Passing (GAMP) is a generalized application of the Approximate Message Passing (AMP) algorithm. Its core idea is to apply quadratic approximation and Taylor technique approximation to cyclic message passing on the basis of the approximate message passing algorithm, and describe the process by a set of simple state evolution equations, so as to reduce the complexity of the algorithm and improve the estimation efficiency. Based on the above theory, by formulating the variational Bayesian (VB) and generalized approximate message propagation into a sparse Bayesian learning model, the reconstruction model level is deepened to make it more in line with the actual problem model, thereby improving the reconstruction quality of the image. At the same time, by increasing the number of levels, the matrix inversion operation in the original problem solving process is transformed, thereby reducing the high computational overhead brought by the matrix inversion, so as to solve large-scale problems in industrial production.

[0004] In order to solve the problems existing in the background technology, the present invention provides a compression and reconstruction method for industrial images, which targets the characteristics of the transmission process of large-scale images produced by industrial production to ensure low-dimensional transmission of industrial image data and high-precision reconstruction after transmission.

[0005] The technical solution adopted by the present invention is:

[0006] The compression and reconstruction method of the present invention comprises the following steps:

[0007] S1. Acquire an industrial image, grayscale the industrial image, and crop it into a number of n×n pixel images to be processed. For each image to be processed, perform sparse feature extraction on the image to be processed by wavelet transform to obtain a sparse image; specifically, the image can be cropped into a number of 512×512 pixel images to be processed.

[0008] S2. In order to restore and reconstruct each sparse image, a restoration model of each sparse image is established, each sparse image is iteratively processed through the restoration model, and a final restoration coefficient of each sparse image is output.

[0009] S3. For each final restoration coefficient, the restoration coefficient is processed by inverse wavelet transform to obtain a restored image.

[0010] The restored images are stitched together to obtain a stitched image, that is, a reconstruction result of the industrial image.

[0011] In the step S2, a restoration model of each sparse image y is established, specifically as follows:

[0012] y=DS(ω)+δ(θ)

[0013] Where D is the dictionary matrix, S(ω) is the restoration coefficient, δ(θ) is the possible linear noise; ω and θ are the first parameter and the second parameter of the restoration model respectively.

[0014] Each sparse image is iteratively processed through the restoration model and the restoration coefficient of each sparse image is output, including the following steps:

[0015] S2.1. The recovery model uses the variational generalized approximate message passing (VGAMP) algorithm to iteratively estimate the updated value of the recovery coefficient of the recovery model, that is, the updated recovery coefficient, and obtains the mean and variance of the updated recovery coefficient through the first parameter and the second parameter preset in the recovery model and the updated recovery coefficient.

[0016] The variational generalized approximate message passing (VGAMP) algorithm is specifically a variational generalized approximate message passing (GAMP) algorithm based on variational Bayes.

[0017] S2.2. According to the mean and variance of the updated recovery coefficient obtained in step S2.1, the first parameter and the second parameter are iteratively updated to obtain the first final updated parameter and the second final updated parameter. The recovery model updates the updated recovery coefficient through the first final updated parameter and the second final updated parameter, and outputs the final recovery coefficient.

[0018] In the step S2.1, the recovery model uses the variational generalized approximate message passing VGAMP algorithm to iteratively estimate the updated value of the recovery coefficient S(ω), that is, the updated recovery coefficient, and obtains the mean value μ of the updated recovery coefficient through the first parameter ω and the second parameter θ preset in the recovery model and the updated recovery coefficient. S and variance Σ S , as follows:

[0019] The recovery model is iteratively estimated by the variational generalized approximate message passing (VGAMP) algorithm until the recovery model is iterated to the jth time, and the recovery coefficient S(ω) is iterated to the jth update value S (j) , and calculate the updated value S (j) Expectations and variance The details are as follows:

[0020]

[0021]

[0022] in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The variance of The update value S of the restoration coefficient S(ω) iterated to the jth time (j) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate variance of is the first regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm used to control the convergence speed; is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate mean of .

[0023] The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate variance of The details are as follows:

[0024]

[0025]

[0026]

[0027]

[0028] in,(·) T Transpose the matrix; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The noise-free model k (j) The approximate mean of (j) =DS (j) ; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate variance of k is the noise-free model (j) The approximate variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The corresponding noise-free model k (j-1) The intermediate variable c (j-1) The approximate mean of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) expectations.

[0029] The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate mean The details are as follows:

[0030]

[0031]

[0032] in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate mean of It is the second regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm used to control the convergence speed.

[0033] The First Regularization Parameter of the Variational Generalized Approximate Message Passing VGAMP Algorithm and the second regularization parameter The details are as follows:

[0034]

[0035] Among them, Σ Λ is the sum of the singular values ​​of the dictionary matrix D, Λ max is the maximum singular value of the dictionary matrix D; α and β are the first constant coefficient and the second constant coefficient respectively; const is the first regularization parameter and the second regularization parameter The constant term of .

[0036] At this time, the recovery coefficient S(ω) iterates to the jth updated value S (j) Expectations Satisfy the following formula:

[0037]

[0038] Among them, δ vgamp is the first cutoff threshold.

[0039] Or at this time, the number of iterations j of the recovery coefficient S(ω) is equal to the first cutoff maximum value J max, then the iterative update is stopped and the iteration of the recovery coefficient S(ω) is completed. At this time, the recovery coefficient S(ω) is iterated to the jth update value S (j) To update the restoration coefficient, according to the first parameter ω and the second parameter θ and the updated restoration coefficient S (j) , the mean value μ of the updated recovery coefficient is obtained S and variance Σ S , that is, the restoration coefficient S(ω) is iterated to the jth update value S (j) Expectations And the recovery coefficient S(ω) is iterated to the jth updated value S (j) Variance Right now

[0040] In step S2.2, the restoration coefficient μ is updated according to the value obtained in step S2.1. S and variance Σ S , iteratively update the first parameter ω and the second parameter θ, obtain the first final update parameter and the second final update parameter, and the recovery model updates the recovery coefficient S through the first final update parameter and the second final update parameter (j) Update and output the final recovery coefficient, as follows:

[0041] The first parameter ω and the second parameter θ are iteratively updated according to the variational Bayesian method. The updated values ​​of the first parameter ω and the second parameter θ are obtained in each iteration, and the first parameter ω and the second parameter θ are directly replaced respectively, thereby updating the updated recovery coefficient S corresponding to the first parameter ω and the second parameter θ. (j) and its mean μ S and variance Σ S .

[0042] That is, while the first parameter ω and the second parameter θ are iterating, the update recovery coefficient S corresponding to the first parameter ω and the second parameter θ (j) and its mean μ S and variance Σ S Iterations are also performed accordingly.

[0043] The updated values ​​of the first parameter ω and the second parameter θ directly replace the updated recovery coefficient S in step S2.1 (j) The first parameter ω and the second parameter θ in each formula are used to obtain the updated restoration coefficient S (j) and its mean μ S and variance Σ S The updated value of .

[0044] Until the first parameter ω and the second parameter θ are iterated to the i-th time, the updated values ​​ω of the first parameter ω and the second parameter θ are obtained respectively (i) and θ (i) , as follows:

[0045]

[0046]

[0047] Among them, a (i) and b (i) are the updated values ​​ω of the first parameter ω updated iteratively to i times (i) The first control parameter and the second control parameter of g (i) and f (i) are the updated values ​​θ of the second parameter θ updated iteratively to i times (i) a first control parameter and a second control parameter; and They are respectively the updated restoration coefficient S (j) Updated value S after iteration i (ji) The mean and variance of ; m is the dimension value of the sparse image y.

[0048] At this time, update the recovery coefficient to the mean of the i-th iteration Satisfy the following formula:

[0049]

[0050]

[0051] Among them, δ sbl is the second cutoff threshold.

[0052] Or at this time, the number of iterations of the first parameter ω and the second parameter θ is greater than the second cutoff value I max ,and Then stop the iterative update and update the first parameter ω at this time to the updated value ω of the i-th iteration (i) As the first final update parameter, the second parameter θ is updated iteratively to the update value θ of the i-th update. (i) As the second final update parameter, the recovery model updates the recovery coefficient S by the first final update parameter and the second final update parameter. (j) Update and output the updated recovery coefficient S at this time (j) Updated value S after iteration i (ji) as the final restitution coefficient.

[0053] If the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Second, I x >I max , never satisfies the first final update parameter ω (j) and the second final update parameter θ (j)The completion condition for stopping the iterative update is, then the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Update value at this time and Substitute into the recovery model, directly replace the first parameter ω and the second parameter θ of the recovery model, and perform the same operation of the recovery model in step S2 on the replaced recovery model, until the final updated value of the first parameter ω and the second parameter θ of the recovery model obtains after several iterations to meet the completion condition for stopping the iterative update of the above-mentioned first parameter ω and the second parameter θ, and outputs the final updated value of the recovery coefficient S(ω) of the recovery model at this time as the final recovery coefficient.

[0054] The beneficial effects of the present invention are:

[0055] The method of the present invention can effectively compress large-scale images in industrial production, improve transmission efficiency, and can quickly and accurately restore the compressed images in the reconstruction after compression for later detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 is the original reference image of the present invention;

[0058] Figure 3 is the image after thinning of the present invention;

[0059] Figure 4 It is the restored image after sparse reconstruction by the method of this paper. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] The specific embodiments are as follows:

[0062] like Figure 1 As shown, the method of the present invention comprises the following steps:

[0063] S1, obtain four industrial images, such as Figure 2 As shown in , each industrial image is grayed and cropped into several n×n pixel images to be processed. For each image to be processed, sparse feature extraction is performed on the image to be processed by wavelet transform to obtain a sparse image, as shown in Figure 3 As shown, there are four sparse images corresponding to four industrial images; specifically, they can be cropped into several 512×512 pixel images to be processed.

[0064] S2. In order to restore and reconstruct each sparse image, a restoration model of each sparse image is established, each sparse image is iteratively processed through the restoration model, and a final restoration coefficient of each sparse image is output.

[0065] In step S2, a restoration model for each sparse image y is established, as follows:

[0066] y=DS(ω)+δ(θ)

[0067] Where D is the dictionary matrix, S(ω) is the recovery coefficient, δ(θ) is the possible linear noise; ω and θ are the first and second parameters of the recovery model respectively; y∈R m×n , D∈R m×n ,S(ω)∈R n×n ,δ(θ)∈R m .

[0068] Each sparse image is iteratively processed through the restoration model and the restoration coefficient of each sparse image is output, including the following steps:

[0069] S2.1. The recovery model uses the variational generalized approximate message passing (VGAMP) algorithm to iteratively estimate the updated value of the recovery coefficient of the recovery model, that is, the updated recovery coefficient, and obtains the mean and variance of the updated recovery coefficient through the first parameter and the second parameter preset in the recovery model and the updated recovery coefficient.

[0070] The variational generalized approximate message passing (VGAMP) algorithm is specifically a variational generalized approximate message passing (GAMP) algorithm based on variational Bayes.

[0071] The recovery model uses the variational generalized approximate message passing (VGAMP) algorithm to iteratively estimate the updated value of the recovery coefficient S(ω), that is, the updated recovery coefficient, and obtains the mean value μ of the updated recovery coefficient through the first parameter ω and the second parameter θ preset in the recovery model and the updated recovery coefficient. S and variance Σ S , as follows:

[0072] The recovery model is iteratively estimated by the variational generalized approximate message passing (VGAMP) algorithm until the recovery model is iterated to the jth time, and the recovery coefficient S(ω) is iterated to the jth update value S (j) , and calculate the updated value S (j) Expectations and variance The details are as follows:

[0073]

[0074]

[0075] in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The variance of The update value S of the restoration coefficient S(ω) iterated to the jth time (j) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate variance of is the first regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm used to control the convergence speed; is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate mean of .

[0076] The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate variance of The details are as follows:

[0077]

[0078]

[0079]

[0080]

[0081] in,(·) T Transpose the matrix; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The noise-free model k (j) The approximate mean of (j) =DS (j) ; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate variance of k is the noise-free model (j) The approximate variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The corresponding noise-free model k (j-1) The intermediate variable c (j-1) The approximate mean of Iterate the recovery coefficient S(ω) to the j-th 1 The updated value S (j-1) The expectation of; the approximate mean of the initial intermediate variable c

[0082] The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate mean The details are as follows:

[0083]

[0084]

[0085] in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate mean of It is the second regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm used to control the convergence speed.

[0086] The First Regularization Parameter of the Variational Generalized Approximate Message Passing VGAMP Algorithm and the second regularization parameter The details are as follows:

[0087]

[0088] Among them, Σ Λ is the sum of the singular values ​​of the dictionary matrix D, Λ max is the maximum singular value of the dictionary matrix D; α and β are the first constant coefficient and the second constant coefficient respectively; const is the first regularization parameter and the second regularization parameter The constant term of .

[0089] At this time, the recovery coefficient S(ω) iterates to the jth updated value S (j) Expectations Satisfy the following formula:

[0090]

[0091] Among them, δ vgamp is the first cutoff threshold, specifically the first cutoff threshold δ vgamp =1e-4.

[0092] Or at this time, the number of iterations j of the recovery coefficient S(ω) is equal to the first cutoff maximum value J max , specifically the first cutoff maximum value J max=100, the iterative update stops and the iteration of the recovery coefficient S(ω) is completed. At this time, the recovery coefficient S(ω) is iterated to the jth update value S (j) To update the restoration coefficient, according to the first parameter ω and the second parameter θ and the updated restoration coefficient S (j) , the mean value μ of the updated recovery coefficient is obtained S and variance Σ S , that is, the restoration coefficient S(ω) is iterated to the jth update value S (j) Expectations And the recovery coefficient S(ω) is iterated to the jth updated value S (j) Variance Right now

[0093] S2.2. According to the mean and variance of the updated recovery coefficient obtained in step S2.1, the first parameter and the second parameter are iteratively updated to obtain the first final updated parameter and the second final updated parameter. The recovery model updates the updated recovery coefficient through the first final updated parameter and the second final updated parameter, and outputs the final recovery coefficient.

[0094] In step S2.2, the restoration coefficient μ is updated according to the value obtained in step S2.1. S and variance Σ S , iteratively update the first parameter ω and the second parameter θ, obtain the first final update parameter and the second final update parameter, and the recovery model updates the recovery coefficient S through the first final update parameter and the second final update parameter (j) Update and output the final recovery coefficient, as follows:

[0095] The first parameter ω and the second parameter θ are iteratively updated according to the variational Bayesian method. The updated values ​​of the first parameter ω and the second parameter θ are obtained in each iteration, and the first parameter ω and the second parameter θ are directly replaced respectively, thereby updating the updated recovery coefficient S corresponding to the first parameter ω and the second parameter θ. (j) and its mean μ S and variance Σ S .

[0096] That is, while the first parameter ω and the second parameter θ are iterating, the update recovery coefficient S corresponding to the first parameter ω and the second parameter θ (j) and its mean μ S and variance Σ S Iterations are also performed accordingly.

[0097] The updated values ​​of the first parameter ω and the second parameter θ directly replace the updated recovery coefficient S in step S2.1 (j) The first parameter ω and the second parameter θ in each formula are used to obtain the updated restoration coefficient S (j) and its mean μS and variance Σ S The updated value of .

[0098] Until the first parameter ω and the second parameter θ are iterated to the i-th time, the updated values ​​ω of the first parameter ω and the second parameter θ are obtained respectively (i) and θ (i) , as follows:

[0099]

[0100]

[0101] Among them, a (i) and b (i) are the updated values ​​ω of the first parameter ω updated iteratively to i times (i) The first control parameter and the second control parameter have initial values ​​of a (1) =b (1) =1e-6; g (i) and f (i) are the second final update parameters θ (j) Update the updated value θ after iteration i (ji) The first control parameter and the second control parameter have initial values ​​of g (1) =f (1) =1e-6; and They are respectively the updated restoration coefficient S (j) Updated value S after iteration i (ji) The mean and variance of ; m is the dimension value of the sparse image y.

[0102] At this time, update the recovery coefficient to the mean of the i-th iteration Satisfy the following formula:

[0103]

[0104]

[0105] Among them, δ sbl is the second cutoff threshold, specifically the second cutoff threshold δ sbl =3e-5.

[0106] Or at this time, the number of iterations of the first parameter ω and the second parameter θ is greater than the second cutoff value I max , specifically the second cutoff value I max =1500, and Then stop the iterative update and update the first parameter ω at this time to the updated value ω of the i-th iteration (i) As the first final update parameter, the second parameter θ is updated iteratively to the update value θ of the i-th update. (i)As the second final update parameter, The recovery model updates the recovery coefficient S by the first final update parameter and the second final update parameter. (j) Update and output the updated recovery coefficient S at this time (j) Updated value S after iteration i (ji) as the final restitution coefficient.

[0107] If the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Second, I x >I max , never satisfies the first final update parameter ω (j) and the second final update parameter θ (j) The completion condition for stopping the iterative update is, then the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Update value at this time and Substitute into the recovery model, directly replace the first parameter ω and the second parameter θ of the recovery model, and perform the same operation of the recovery model in step S2 on the replaced recovery model, until the final updated value of the first parameter ω and the second parameter θ of the recovery model obtains after several iterations to meet the completion condition for stopping the iterative update of the above-mentioned first parameter ω and the second parameter θ, and outputs the final updated value of the recovery coefficient S(ω) of the recovery model at this time as the final recovery coefficient.

[0108] S3. For each final restoration coefficient, the restoration coefficient is processed by inverse wavelet transform to obtain a restored image; each restored image is spliced ​​to obtain a spliced ​​image, that is, the reconstruction result of the industrial image is obtained, such as Figure 4 shown.

[0109] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope of the present invention.

Claims

1. A compression and reconstruction method for industrial images, characterized in that: The steps include: S1. Acquire an industrial image, grayscale the industrial image, and cut it into a number of n×n pixel images to be processed. For each image to be processed, perform sparse feature extraction on the image to be processed by wavelet transform to obtain a sparse image; S2, establishing a restoration model for each sparse image, iteratively processing each sparse image through the restoration model and outputting a final restoration coefficient of each sparse image; S3, for each final restoration coefficient, the restoration coefficient is processed by inverse wavelet transform to obtain a restored image; The restored images are stitched together to obtain a stitched image, that is, a reconstruction result of the industrial image; In the step S2, a restoration model of each sparse image y is established, specifically as follows: y=DS(ω)+δ(θ) Where D is the dictionary matrix, S(ω) is the recovery coefficient, δ(θ) is the linear noise; ω and θ are the first parameter and the second parameter of the recovery model respectively; Each sparse image is iteratively processed through the restoration model and the restoration coefficient of each sparse image is output, including the following steps: S2.1, the recovery model uses the variational generalized approximate message passing (VGAMP) algorithm to iteratively estimate the updated value of the recovery coefficient of the recovery model, that is, the updated recovery coefficient, and obtains the mean and variance of the updated recovery coefficient through the first parameter and the second parameter preset in the recovery model and the updated recovery coefficient; S2.2, according to the mean and variance of the updated recovery coefficient obtained in step S2.1, iteratively update the first parameter and the second parameter to obtain the first final updated parameter and the second final updated parameter, and the recovery model updates the updated recovery coefficient by the first final updated parameter and the second final updated parameter, and outputs the final recovery coefficient; In the step S2.1, the recovery model uses the variational generalized approximate message passing VGAMP algorithm to iteratively estimate the updated value of the recovery coefficient S(ω), that is, the updated recovery coefficient, and obtains the mean value μ of the updated recovery coefficient through the first parameter ω and the second parameter θ preset in the recovery model and the updated recovery coefficient. S and variance∑ S , as follows: The recovery model is iteratively estimated by the variational generalized approximate message passing (VGAMP) algorithm until the recovery model is iterated to the jth time, and the recovery coefficient S(ω) is iterated to the jth update value S (j) , and calculate the updated value S (j) Expectations and variance The details are as follows: in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The variance of The update value S of the restoration coefficient S(ω) iterated to the jth time (j) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate variance of is the first regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm; is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) expectations; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The approximate mean of The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate variance of The details are as follows: in,(·) T Transpose the matrix; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) The noise-free model k (j) The approximate mean of (j) =DS (j) ; The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate variance of k is the noise-free model (j) The approximate variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The variance of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) The corresponding noise-free model k (j-1) The intermediate variable c (j-1) The approximate mean of is the update value S of the restoration coefficient S(ω) iterated to the j-1th time (j-1) expectations; The restoration coefficient S(ω) is iterated to the jth updated value S (j) The approximate mean The details are as follows: in, The update value S of the restoration coefficient S(ω) iterated to the jth time (j) With the noise-free model k (j) The intermediate variable c (j) The approximate mean of is the second regularization parameter of the variational generalized approximate message passing (VGAMP) algorithm; The First Regularization Parameter of the Variational Generalized Approximate Message Passing VGAMP Algorithm and the second regularization parameter The details are as follows: Among them, Σ Λ is the sum of the singular values ​​of the dictionary matrix D, Λ max is the maximum singular value of the dictionary matrix D; α and β are the first constant coefficient and the second constant coefficient respectively; const is the first regularization parameter η k and the second regularization parameter The constant term of ; At this time, the recovery coefficient S(ω) iterates to the jth updated value S (j) Expectations Satisfy the following formula: Among them, δ vgamp is the first cutoff threshold; Or at this time, the number of iterations j of the recovery coefficient S(ω) is equal to the first cutoff maximum value J max , then the iterative update is stopped and the iteration of the recovery coefficient S(ω) is completed. At this time, the recovery coefficient S(ω) is iterated to the jth update value S (j) To update the restoration coefficient, according to the first parameter ω and the second parameter θ and the updated restoration coefficient S (j) , the mean value μ of the updated recovery coefficient is obtained S and variance∑ S , that is, the restoration coefficient S(ω) is iterated to the jth update value S (j) Expectations And the recovery coefficient S(ω) is iterated to the jth updated value S (j) Variance Right now 2. The method for compressing and reconstructing industrial images according to claim 1, characterized in that: In step S2.2, the restoration coefficient μ is updated according to the value obtained in step S2.

1. S and variance Σ S , iteratively update the first parameter ω and the second parameter θ, obtain the first final update parameter and the second final update parameter, and the recovery model updates the recovery coefficient S through the first final update parameter and the second final update parameter (j) Update and output the final recovery coefficient, as follows: The first parameter ω and the second parameter θ are iteratively updated according to the variational Bayesian method. The updated values ​​of the first parameter ω and the second parameter θ are obtained in each iteration, and the first parameter ω and the second parameter θ are directly replaced respectively, thereby updating the updated recovery coefficient S corresponding to the first parameter ω and the second parameter θ. (j) and its mean μ S and variance∑ S ; Until the first parameter ω and the second parameter θ are iterated to the i-th time, the updated values ​​ω of the first parameter ω and the second parameter θ are obtained respectively (i) and θ (i) , as follows: Among them, a (i) and b (i) are the updated values ​​ω of the first parameter ω updated iteratively to i times (i) The first control parameter and the second control parameter of g (i) and f (i) are the updated values ​​θ of the second parameter θ updated iteratively to i times (i) a first control parameter and a second control parameter; and They are respectively the updated restoration coefficient S (j) Updated value S after iteration i (ji) The mean and variance of ; m is the dimension value of the sparse image y; At this time, update the recovery coefficient to the mean of the i-th iteration Satisfy the following formula: Among them, δ sbl is the second cutoff threshold; Or at this time, the number of iterations of the first parameter ω and the second parameter θ is greater than the second cutoff value I max ,and Then stop the iterative update and update the first parameter ω at this time to the updated value ω of the i-th iteration (i) As the first final update parameter, the second parameter θ is updated iteratively to the update value θ of the i-th update. (i) As the second final update parameter, the recovery model updates the recovery coefficient S by the first final update parameter and the second final update parameter. (j) Update and output the updated recovery coefficient S at this time (j) Updated value S after iteration i (ji) As the final restitution coefficient; If the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Second, I x >I max , never satisfies the first final update parameter ω (j) and the second final update parameter θ (j) The completion condition for stopping the iterative update is, then the first parameter ω and the second parameter θ are iterated to the third cutoff value I x Update value at this time and Substitute into the recovery model, directly replace the first parameter ω and the second parameter θ of the recovery model, and perform the same operation of the recovery model in step S2 on the replaced recovery model, until the final updated value of the first parameter ω and the second parameter θ of the recovery model obtains after several iterations to meet the completion condition for stopping the iterative update of the above-mentioned first parameter ω and the second parameter θ, and outputs the final updated value of the recovery coefficient S(ω) of the recovery model at this time as the final recovery coefficient.

Images