Self-adaptive block compressed sensing image super-resolution reconstruction method and system

Abstract

The invention discloses an adaptive blocking compressed sensing image super-resolution reconstruction method and system. The method comprises the following steps: acquiring a low-resolution image and carrying out adaptive blocking on the low-resolution image; establishing a sparse representation model of the high-resolution image by using a discrete cosine transform basis; establishing an observation dimension reduction model by using the binary block diagonal matrix to obtain a relationship between a high-resolution image and a low-resolution image; performing compressed sensing super-resolution reconstruction solution on the low-resolution sub-block images by using a weighted iterative least square algorithm to generate high-resolution image blocks, and performing constraint by using Shearlet transformation and dual-scale total variation regularization in the reconstruction process; and splicing the generated high-resolution image blocks according to the position indexes to obtain a complete high-resolution image meeting the requirements. According to the method, the problems of unreasonable blocking caused by single blocking basis and limited reconstruction effect caused by insufficient consideration of prior information in the existing method are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image super-resolution reconstruction, and relates to an adaptive block-based compressed sensing image super-resolution reconstruction method and system. Background Art

[0002] In the fields of medicine, security, remote sensing, etc., high-quality, detail-rich images are essential for accurate identification, monitoring and analysis. However, due to factors such as environmental interference, equipment limitations and signal transmission loss, the collected images often have insufficient resolution. How to effectively improve image clarity based on low-resolution images and overcome the challenges brought by low-quality images has become a technical problem that needs to be solved urgently in various industries. Image super-resolution technology uses digital signal processing methods to reconstruct low-resolution images into high-resolution images, explore the potential information value of images without increasing hardware costs, and break through the resolution limitations of equipment. Therefore, it has broad prospects in practical applications.

[0003] At present, single image super-resolution methods mainly include interpolation, reconstruction and deep learning-based algorithms. Interpolation methods (such as bilinear interpolation and cubic spline interpolation) are simple to calculate, but because they cannot accurately model complex image information, their reconstructed images often appear over-smoothed or over-sharpened. Deep learning methods (such as convolutional neural networks and generative adversarial networks) perform well in reconstruction quality, but have high requirements on the diversity and scale of training data sets. At the same time, their computational cost is high, and the training process requires a lot of computing resources, which also limits their feasibility in real-time applications.

[0004] With the development of compressed sensing (CS) theory, it has been widely used in the field of super-resolution reconstruction. This method uses the sparse characteristics of images in a specific transform domain to achieve high-quality reconstruction with limited observation data. However, the traditional compressed sensing super-resolution method still has the following problems:

[0005] First, the basic compressed sensing reconstruction method is difficult to effectively combine the prior information of the image, which limits its ability to restore high-frequency texture details, thus affecting the quality of super-resolution reconstruction. Secondly, the method has high computational and storage overheads. The dimension of the observation matrix corresponding to large-scale images increases dramatically, which not only increases the computational complexity, but also puts higher requirements on storage resources, seriously restricting its practical application in large-scale data scenarios. In order to reduce computational and storage overheads, images are usually processed in blocks, but the compressed sensing reconstruction method of uniform block division is prone to obvious block effects. At the same time, using a single parameter as the basis for segmentation is difficult to fully consider the texture richness of the image, resulting in information loss or over-smoothing. Therefore, there is an urgent need for an efficient adaptive block compressed sensing super-resolution reconstruction method to improve the reconstruction quality while ensuring computational efficiency.

[0006] To solve the above problems, the present invention proposes a compressive sensing image super-resolution reconstruction method combining adaptive block technology, Shearlet transform and dual-scale total variation regularization. Summary of the Invention

[0007] The purpose of the present invention is to provide an adaptive block-based compressive sensing image super-resolution reconstruction method and system, which solves the problems of unreasonable block division caused by single block division basis and limited reconstruction effect caused by insufficient consideration of prior information in the existing adaptive block compressive sensing super-resolution method.

[0008] The technical solution for realizing the purpose of the present invention is as follows:

[0009] An adaptive block-based compressive sensing image super-resolution reconstruction method, comprising:

[0010] Step 1, obtaining a low-resolution image and performing adaptive block division on it;

[0011] Step 2, based on the set requirements of the high-resolution image, establishing a sparse representation model of the high-resolution image using the discrete cosine transform basis;

[0012] Step 3, based on the sparse representation model of the high-resolution image, establishing an observation dimensionality reduction model using a binary block diagonal matrix to obtain the relationship between the high-resolution image and the low-resolution image;

[0013] Step 4, based on the observation dimensionality reduction model, using the weighted iterative least squares algorithm to solve the compressive sensing super-resolution reconstruction of the low-resolution sub-block images, generating high-resolution image blocks, and using Shearlet transform and dual-scale total variation regularization for constraint during the reconstruction process;

[0014] Step 5, splicing the generated high-resolution image blocks according to the position index to obtain a complete high-resolution image that meets the requirements.

[0015] Further, the process of obtaining the low-resolution image and performing adaptive block division on it includes: uniformly dividing the input image into four initial sub-blocks, calculating the comprehensive entropy of each sub-block, and taking the average value of the comprehensive entropy of all sub-blocks as the segmentation threshold τ. Recursively divide the sub-blocks whose comprehensive entropy is greater than the average value τ and greater than the minimum size, and stop dividing the sub-blocks whose comprehensive entropy is less than the average value τ or the size is insufficient, and then output the adaptive block division result that meets the termination conditions.

[0016] Further, the comprehensive entropy E combined is calculated by the formula:

[0017]

[0018] Among them, α is the weight coefficient of gray entropy, β is the weight coefficient of the contrast of the Gray Level Co-occurrence Matrix (GLCM), and α + β = 1. μ and σ represent the mean and standard deviation respectively. The gray entropy E Entropy is calculated as follows:

[0019]

[0020] The GLCM contrast E Contrast is calculated as follows:

[0021]

[0022] where P(i, j) is the normalized probability distribution of the gray level co-occurrence matrix of the image, i and j represent the gray levels of the image respectively, and P(i, j) represents the joint occurrence probability distribution of the pixel gray values i and j at a given spatial distance and direction, which is used to characterize the gray level co-occurrence relationship of the image and further used to calculate the entropy value and contrast of the image to measure the texture features of the image.

[0023] Furthermore, a sparse representation model of the high-resolution image is established using the discrete cosine transform basis, which specifically includes:

[0024] The high-resolution image X ∈ R N×N is sparsely represented on the DCT basis Ψ: where Ψ is the DCT transform basis, is the sparse coefficient vector in the DCT domain.

[0025] Furthermore, the process of establishing the observation dimensionality reduction model includes:

[0026] For the high-resolution image X ∈ R N×N , first, the columns of the high-resolution image are compressed by using a binary block diagonal matrix as the measurement matrix Φ ∈ R M×N (M < N) to obtain an intermediate result Z ∈ R M×N . Subsequently, the intermediate result Z ∈ R M×N is transposed and compressed again by the measurement matrix along the row direction to obtain the final low-resolution image Y ∈ R M×M . The overall compression process is expressed as Equation Y M×M = Φ M×N ·(Φ M×N ·X N×N ) T . (·) T represents the transpose operation.

[0027] Furthermore, the process of using the Shearlet transform and double-scale total variation regularization for constraint includes:

[0028] Shearlet transform regularization:

[0029] In the Shearlet transform, the input image is projected onto a set of multi-scale and multi-directional transform basis functions to extract edge and texture details. The transform kernel is composed of the shear matrix B s and the anisotropic dilation matrix A a generated as follows:

[0030]

[0031] where s represents the shear parameter, which is used to capture the directional features of the image. a represents the scale parameter, which is used to control the scale change of the basis functions. Based on these two matrices, the shearlet transform kernel can be defined and represented by where x represents the two-dimensional signal, and t is the translation parameter used to control the position. The definition of the shearlet transform of the input function f(x) under this transform kernel is as shown in Equation as follows.

[0032] Dual-scale total variation regularization:

[0033] The l 1 norm of the image gradient is constrained by adopting dual-scale total variation regularization. At the same time, the first-scale gradient TV 1 and the second-scale gradient TV 2 of the image are calculated to achieve the comprehensive constraint of gradient information at different scales. Suppose the pixel value at the i-th row and j-th column of the image is x i,j , then their calculation formulas are respectively as shown in Equation and The total variation regularization terms at the two scales are linearly combined to form a comprehensive constraint term, and its expression is as follows:

[0034] DTV(x) = α 1 ·TV 1 (x) + α 2 ·TV 2 (x).

[0035] Furthermore, the compressive sensing super-resolution reconstruction is solved for the low-resolution sub-block images by using the weighted iterative least squares algorithm, which specifically includes:

[0036] Using the Shearlet transform and dual-scale total variation regularization to construct an objective function for optimizing super-resolution reconstruction;

[0037] Introducing two auxiliary variables corresponding to the Shearlet transform regularization term and the dual-scale total variation regularization term respectively, and substituting them into the objective function;

[0038] The Lagrange multiplier is introduced into the objective function to construct an augmented Lagrangian function, and the alternating direction method of multipliers is used to solve it to obtain the high-resolution image block. In each iteration, the weight matrix is dynamically updated.

[0039] Furthermore, the objective function for super-resolution reconstruction optimization is constructed as:

[0040]

[0041] where ||·|| 1 represents the 1-norm, y represents the low-resolution image, is the sparse coefficient of X in the sparse basis Ψ, Φ is the observation matrix, and λ 1 and λ 2 are constant coefficients. The fidelity term in the objective function is used to ensure that the reconstruction result is consistent with the observed data, represents the Shearlet transform sparse regularization term, i represents the number of sub-bands, represents the double-scale total variation regularization term.

[0042] Furthermore, by introducing auxiliary variables z 1 and z 2 corresponding to the Shearlet transform regularization term and the DTV regularization term respectively, the problem is rewritten as:

[0043] Furthermore, the Lagrange multiplier is introduced into the objective function to construct an augmented Lagrangian function, and the alternating direction method of multipliers is used to solve it to obtain the high-resolution image block. In each iteration, the weight matrix is dynamically updated, specifically including: introducing Lagrange multipliers μ 1 、μ 2 , constructing the augmented Lagrangian function:

[0044]

[0045] The objective function is optimized and solved using the alternating direction method of multipliers. In each iteration, according to the currently reconstructed the weight matrix is dynamically updated where ∈ is a very small positive number used to avoid a zero denominator. For expand by rows, and the formula is:

[0046] A row =Ψ T φ T W (k) φΨ+ρ 2 I+ρ 1 SH T SH

[0047]

[0048] For expansion by column, the formula is:

[0049] A col = Ψ T φ T W (k) φΨ + ρ 2 I + ρ 1 SH T SH

[0050]

[0051] For the Shearlet regularization problem with respect to z 1 solve:

[0052]

[0053] For the DTV regularization problem with respect to z 2 solve:

[0054]

[0055] Update the Lagrange multipliers.

[0056]

[0057] Through the above optimization process, for the variables z 1 、z 2 and the Lagrange multipliers μ 1 、μ 2 alternatively solve to achieve the reconstruction of high - resolution images Thus, the efficient and accurate super - resolution reconstruction of the input low - resolution image is achieved.

[0058] An adaptive block - based compressive sensing image super - resolution reconstruction system, comprising:

[0059] An image partitioning module, configured to obtain a low - resolution image and adaptively partition it;

[0060] A sparse representation model construction unit, based on the set requirements of high - resolution images, uses the discrete cosine transform basis to establish a sparse representation model of high - resolution images;

[0061] An observation dimensionality reduction model construction unit, based on the sparse representation model of high - resolution images, uses a binary block - diagonal matrix to establish an observation dimensionality reduction model to obtain the relationship between high - resolution images and low - resolution images;

[0062] A solution unit, based on an observation dimensionality reduction model, uses a weighted iterative least squares algorithm to perform compressive sensing super-resolution reconstruction on a low-resolution sub-block image, generating a high-resolution image block. During the reconstruction process, the Shearlet transform and double-scale total variation regularization are used for constraint;

[0063] A stitching unit stitches the generated high-resolution image blocks according to the position index to obtain a complete high-resolution image that meets the requirements.

[0064] Compared with the prior art, the significant advantages of the present invention are:

[0065] (1) In the adaptive block division process of the present invention, the texture richness is comprehensively characterized by the gray entropy and GLCM contrast, and this is used as the basis for adaptive block division, which helps to comprehensively and accurately capture the texture characteristics of different regions. At the same time, the block division process effectively reduces the scale of the observation matrix and the single calculation complexity, alleviating the dimensionality disaster problem in compressive sensing reconstruction.

[0066] (2) In the compressive sensing reconstruction process of the present invention, a binary block diagonal matrix is used as the measurement matrix. This matrix can effectively maintain the structural similarity between the low-resolution image and the high-resolution image, and shows good mutual irrelevance with the DCT sparse basis, and is also easier to implement in hardware.

[0067] (3) The present invention innovatively uses the IRLS algorithm for super-resolution reconstruction, and effectively combines the multi-scale characteristics of the Shearlet transform and the denoising and edge-preserving capabilities of double-scale total variation regularization. By efficiently extracting texture information and enhancing details, suppressing noise and keeping edges clear, the quality of super-resolution reconstruction is significantly improved. Description of the Drawings

[0068] Figure 1 It is a flowchart of the adaptive block division of the input image based on the comprehensive entropy parameter.

[0069] Figure 2 It is a schematic diagram of the block division result of the adaptive block division of the remote sensing image.

[0070] Figure 3 It is a representation diagram of the DCT sparse basis at different frequencies.

[0071] Figure 4 It is a schematic diagram of the binary block diagonal observation matrix.

[0072] Figure 5 It is a flowchart of the adaptive block division compressive sensing image super-resolution.

[0073] Figure 6 It is a schematic diagram of the reconstruction effect of the super-resolution reconstruction of the low-resolution remote sensing image.Figure 6 In (a), it is the input low-resolution image, and in (b), it is the output high-resolution image. Specific implementation mode

[0074] Combined with Figures 1 to 5 , a remote sensing image compressive sensing super-resolution reconstruction method based on adaptive block combination of Shearlet transform and double-scale total variation regularization proposed by the present invention is described in detail.

[0075] The image super-resolution method in this embodiment includes the following contents.

[0076] Step 1: Obtain a low-resolution remote sensing image and perform adaptive block division on it. Combined with Figure 1 and Figure 2 , it specifically includes:

[0077] 1.1 Define the comprehensive entropy through the gray entropy and the GLCM contrast:

[0078] In image texture analysis, the GLCM contrast and the gray entropy are commonly used feature parameters. The GLCM contrast statistics the gray differences of pixel pairs in different directions and distances, focusing on capturing local intensity changes and geometric structure information; the gray entropy is based on the joint probability of the gray distribution and measures the uncertainty and randomness of the image information distribution. Each has its own focus in texture analysis and can complementarily represent the texture characteristics of the image.

[0079] In order to improve the accuracy and robustness of the adaptive block division, the standardized values of the GLCM contrast and the gray entropy are linearly weighted to define the comprehensive entropy. Its calculation formula is:

[0080]

[0081] where α is the weight coefficient of the gray entropy, β is the weight coefficient of the GLCM contrast, α + β = 1, here α is selected as 0.55, β is selected as 0.45. μ and σ respectively represent the mean and the standard deviation. The calculation method of the gray entropy E Entropy is:

[0082]

[0083] The calculation method of the GLCM contrast E Contrast is:

[0084]

[0085] Where P(i, j) is the normalized probability distribution of the gray-level co-occurrence matrix of the image. i and j represent the gray levels of the image respectively. P(i, j) represents the joint occurrence probability distribution of pixel gray values i and j at a given spatial distance and direction, which is used to characterize the gray-level co-occurrence relationship of the image, and further used to calculate the entropy value and contrast of the image to measure the texture features of the image.

[0086] 1.2 Adaptive Blocking Method Based on the Comprehensive Entropy Segmentation Threshold:

[0087] The low-resolution remote sensing image with a size of 256×256 is evenly divided into four initial sub-blocks with a size of 64×64, and the comprehensive entropy value of each sub-block is calculated; the average value of the comprehensive entropy of all sub-blocks is taken as the segmentation threshold τ; if the comprehensive entropy of the sub-block is higher than the threshold τ and the size is larger than the set minimum limit, the sub-block is further subdivided. If the comprehensive entropy is lower than the threshold τ or the sub-block size is less than or equal to 16×16, the segmentation stops; when all sub-blocks meet the termination conditions, the final adaptive blocking result is output.

[0088] This method can achieve fine segmentation of the texture area while ensuring the image structure information, and improve the adaptability and accuracy of blocking.

[0089] Step 2: The process of establishing the sparse representation model using the DCT sparse basis includes:

[0090] Sparse representation of the high-resolution image in the transform domain is performed to facilitate the processing of the super-resolution task. Let the high-resolution image X ∈ R N×N It can be sparsely represented on the DCT basis Ψ: where Ψ is the DCT transform basis, is the sparse coefficient vector in the DCT domain. Since DCT can efficiently decompose the spatial frequency components of the image and achieve sparse representation in the transform domain, it is widely used in the super-resolution image reconstruction task, providing theoretical support for image compression and reconstruction.

[0091] Step 3: Establish the observation dimensionality reduction model using the binary block diagonal matrix:

[0092] To infer the high-resolution image from the low-resolution image, this step establishes an observation dimensionality reduction model from the high-resolution image to the low-resolution image. The high-dimensional data is compressed and mapped through the binary block diagonal matrix as the measurement matrix, providing a theoretical basis for subsequent super-resolution reconstruction.

[0093] For the high-resolution image X ∈ R N×N , first, the columns of the high-resolution image are compressed through the binary block diagonal matrix as the measurement matrix Φ ∈ R M×N (M < N) to obtain the intermediate result Z ∈ R M×N , Z M×N= Φ M×N ·X N×N Subsequently, the intermediate result Z ∈ R M×N is transposed and compressed again by the measurement matrix along the row direction to obtain the final low-resolution image Y ∈ R M ×M .

[0094] The overall compression process is expressed as Equation Y M×M = Φ M×N · [Φ M×N ·X N×N T .

[0095] Step 4: Design the constraints using Shearlet transform and dual-scale total variation regularization:

[0096] In the Shearlet transform, the input image is projected onto a set of multi-scale and multi-directional transform basis functions to extract edge and texture details. The transform kernel is generated by the shear matrix B s and the anisotropic dilation matrix A a as follows:

[0097]

[0098] where s represents the shear parameter for capturing the directional features of the image. a represents the scale parameter for controlling the scale change of the basis functions. Based on these two matrices, the shearlet transform kernel can be defined and represented by . Where x represents the two-dimensional signal and t is the translation parameter for controlling the position. The definition of the shearlet transform of the input function f(x) under this transform kernel is as shown in Equation .

[0099] The l 1 norm of the image gradient is constrained by adopting dual-scale total variation regularization. At the same time, the first-scale gradient TV 1 and the second-scale gradient TV 2 of the image are calculated to achieve the comprehensive constraint of gradient information at different scales. Suppose the pixel value at the i-th row and j-th column of the image is x i,j , then their calculation formulas are as shown in Equation and respectively. The two-scale total variation regularization terms are linearly combined to form a comprehensive constraint term, and its expression is as follows:

[0100] DTV(x) = α 1 ·TV 1 (x) + α 2 ·TV 2 (x).

[0101] ​Step 5. Use the improved weighted iterative least squares algorithm to perform compressive sensing super-resolution reconstruction on the low-resolution sub-block images, and use the Shearlet transform and dual-scale total variation regularization for constraint during the reconstruction process. Specifically, it includes:

[0102] Model the super-resolution reconstruction optimization problem as:

[0103]

[0104] where ||·|| 1 represents the 1-norm, y represents the low-resolution image, is the sparse coefficient of X on the sparse basis Ψ, Φ is the observation matrix, λ 1 and λ 2 are constant coefficients, which are set to 3.00 and 0.06 respectively. The fidelity term in the objective function is used to ensure that the reconstruction result is consistent with the observed data, represents the Shearlet transform sparse regularization term, i represents the number of sub-bands, represents the dual-scale total variation regularization term.

[0105] For the super-resolution reconstruction model to be solved, both the Shearlet transform and the DTV regularization term are non-linear operations, which will lead to a highly coupled and difficult-to-solve optimization problem. Therefore, by introducing auxiliary variables z 1 and z 2 corresponding to the Shearlet transform regularization term and the DTV regularization term respectively, rewrite the problem as:

[0106]

[0107] By introducing Lagrange multipliers μ 1 and μ 2 , construct the augmented Lagrangian function

[0108]

[0109] Use the alternating direction method of multipliers to optimize and solve the objective function. In each iteration, dynamically update the weight matrix according to the currently reconstructed where ∈ is a very small positive number used to avoid a zero denominator. The formula for row-wise expansion of the reconstruction is:

[0110] A row = Ψ T φ T W (k) φΨ + ρ 2 I + ρ 1 SH T SH

[0111]

[0112] The formula for column - expanded reconstruction is as follows:

[0113] A col = Ψ T φ T W (k) φΨ + ρ 2 I + ρ 1 SH T SH

[0114]

[0115] Solve the Shearlet regularization problem with respect to z 1 :

[0116]

[0117] Solve the DTV regularization problem with respect to z 2 :

[0118]

[0119] Finally, update the Lagrange multipliers.

[0120]

[0121] Through the above - mentioned optimization process, alternately solve for the variables z 1 、z 2 and the Lagrange multipliers μ 1 、μ 2 to achieve the reconstruction of high - resolution images This decomposition method significantly reduces the computational complexity of each step of optimization. In addition, the strategy of alternating optimization helps to improve the convergence speed of the algorithm and ensures that the reconstruction quality is gradually improved during the iteration process.

[0122] Step 6: Stitch the generated high - resolution image patches according to the position index to obtain a complete high - resolution remote - sensing image. For the input low - resolution remote - sensing image with a size of 256×256, its reconstruction result is 512×512 in size, achieving a good super - resolution reconstruction effect. The reconstruction result is as Figure 6 shown, and the corresponding PSNR and SSIM values are 24.95dB and 0.7699 respectively, which are improved by 2.82dB and 0.1018 compared with the classic bicubic interpolation method.

[0123] This embodiment also provides a compressive sensing image super-resolution reconstruction system with adaptive block division, including:

[0124] An image division module, configured to obtain a low-resolution image and perform adaptive block division on it;

[0125] A sparse representation model construction unit, based on the set requirements for high-resolution images, uses the discrete cosine transform basis to establish a sparse representation model for high-resolution images;

[0126] An observation dimensionality reduction model construction unit, based on the sparse representation model of high-resolution images, uses a binary block diagonal matrix to establish an observation dimensionality reduction model to obtain the relationship between high-resolution images and low-resolution images;

[0127] A solution unit, based on the observation dimensionality reduction model, uses the weighted iterative least squares algorithm to solve the compressive sensing super-resolution reconstruction of the low-resolution sub-block images, generates high-resolution image blocks, and uses the Shearlet transform and double-scale total variation regularization for constraint during the reconstruction process;

[0128] A splicing unit, splices the generated high-resolution image blocks according to the position index to obtain a complete high-resolution image that meets the requirements.

[0129] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. An adaptive block-based compressed sensing image super-resolution reconstruction method, characterized in that: include: Step 1: Obtain a low-resolution image and adaptively divide it into blocks; Step 2: Based on the set high-resolution image requirements, a sparse representation model of the high-resolution image is established using a discrete cosine transform basis; Step 3: Based on the sparse representation model of the high-resolution image, a binary block diagonal matrix is ​​used to establish an observation dimension reduction model to obtain the relationship between the high-resolution image and the low-resolution image; Step 4: Based on the observation dimension reduction model, the weighted iterative least squares algorithm is used to perform compressed sensing super-resolution reconstruction on the low-resolution sub-block image to generate a high-resolution image block. The Shearlet transform and dual-scale total variation regularization are used for constraints during the reconstruction process. Step 5: splice the generated high-resolution image blocks according to the position index to obtain a complete high-resolution image that meets the requirements.

2. The method for super-resolution reconstruction of compressed sensing images based on adaptive block partitioning according to claim 1, characterized in that: The method for adaptive blocking is as follows: the input image is evenly divided into four initial sub-blocks, the comprehensive entropy of each sub-block is calculated, and the average comprehensive entropy of all sub-blocks is taken as the segmentation threshold, and the sub-blocks whose comprehensive entropy is greater than the segmentation threshold and greater than the minimum size are recursively segmented, and the sub-blocks whose comprehensive entropy is less than the segmentation threshold or the size is insufficient are stopped, and then the adaptive blocking results that meet the termination conditions are output.

3. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 2, characterized in that: The comprehensive entropy is: Among them, α is the weight coefficient of gray entropy, β is the weight coefficient of gray co-occurrence matrix contrast, α+β=1, μ and σ represent the mean and standard deviation respectively, E Entropy is the grayscale entropy, E Contrast is the gray-level co-occurrence matrix contrast.

4. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 3, characterized in that: The grayscale entropy is: Among them, P(i,j) is the standardized probability distribution of the gray-level co-occurrence matrix of the image, i and j represent the gray levels of the image, and P(i,j) represents the joint occurrence probability distribution of pixel gray values ​​i and j in a given spatial distance and direction.

5. The method for super-resolution reconstruction of compressed sensing images based on adaptive block partitioning according to claim 4, characterized in that: The gray-level co-occurrence matrix contrast is:

6. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 1, characterized in that: The compressed sensing super-resolution reconstruction of low-resolution sub-block images using the weighted iterative least squares algorithm specifically includes: The objective function of super-resolution reconstruction optimization is constructed by using Shearlet transform and dual-scale total variation regularization. Two auxiliary variables are introduced, corresponding to the Shearlet transform regularization term and the dual-scale total variation regularization term, and substituted into the objective function; The Lagrangian multiplier is introduced into the objective function, and the augmented Lagrangian function is constructed. The alternating direction multiplier method is used to solve it to obtain the high-resolution image block. In each iteration, the weight matrix is ​​dynamically updated.

7. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 6, characterized in that: The objective function is: Among them, ||·||1 represents the 1-norm, y represents the low-resolution image, is the sparse coefficient of X on the sparse basis Ψ, Φ is the measurement matrix, λ1 and λ2 are constant coefficients, represents the Shearlet transform sparse regularization term, i represents the number of subbands, represents the dual-scale total variation regularization term.

8. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 6, characterized in that: The regularization term using Shearlet transform regularization is: Among them, s represents the shear parameter, a represents the scale parameter, A a and B s are the shear matrix and anisotropic dilation matrix respectively, x represents a two-dimensional signal, t is the translation parameter used to control the position, and f(x) is the input function; The regularization term of dual-scale total variation regularization is: DTV(x)=α1·TV1(x)+α2·TV2(x) Among them, TV1 and TV2 are the first scale gradient and the second scale gradient of the image respectively, and α1 and α2 are coefficients.

9. The method for adaptive block-based compressed sensing image super-resolution reconstruction according to claim 6, characterized in that: The weight matrix is: in, is the solution value of the current iteration.

10. A compressed sensing image super-resolution reconstruction system implementing any of the methods described in claims 1-9, characterized in that: include: An image segmentation module, used to obtain low-resolution images and adaptively segment them; A sparse representation model building unit, based on the set high-resolution image requirements, uses a discrete cosine transform basis to build a sparse representation model of the high-resolution image; The observation dimension reduction model building unit uses a binary block diagonal matrix to build an observation dimension reduction model based on a sparse representation model of a high-resolution image, and obtains the relationship between the high-resolution image and the low-resolution image; The solving unit uses a weighted iterative least squares algorithm to solve the compressed sensing super-resolution reconstruction of low-resolution sub-block images based on the observation dimension reduction model to generate high-resolution image blocks. Shearlet transform and dual-scale total variation regularization are used for constraints during the reconstruction process. The stitching unit stitches the generated high-resolution image blocks according to the position indexes to obtain a complete high-resolution image that meets the requirements.

Images