CN109448065A - A kind of compression sensing method based on the measurement of gradient block adaptive - Google Patents

Abstract

The invention relates to a compressed sensing method based on gradient block adaptive measurement. The method comprises: using a non-overlapping basic block as a dividing unit of an image, calculating a smoothness of each basic block by using a vertical direction and a horizontal gradient of the image, and dividing the image into image blocks of uneven size according to the smoothness and calculating The measurement rate of each image block; the measurement matrix is selected according to the image block size to measure each image block to obtain the measured value; the decoding end introduces the vector angle as the similarity judgment standard, and adopts the non-local low rank regularized compressed sensing reconstruction. The algorithm performs reconstruction. The invention designs a non-uniform block adaptive measurement compressed sensing method, which makes the decoded image robust and obtains better reconstruction effect.

Description

A Compressed Sensing Method Based on Gradient Block Adaptive Measurement

Technical field

The present invention relates to the field of computer image processing, and in particular to a compressed sensing method based on gradient block adaptive measurement.

Background technique

Compressed sensing is a new signal processing framework that breaks through the Nyquist sampling theorem. Information compression and signal reconstruction are two important components of compressed sensing. The information compression methods are mainly divided into two categories, which compress the entire image and compress the image separately. The overall compressed sensing often needs to store a large measurement matrix, occupying a large memory, and at the same time, the calculation of the overall compressed sensing is also very large.

To this end, researchers have proposed a block compression method, which first divides the image into small blocks of a specified size, and then compresses each small block image using the same measurement matrix. For example, the Chinese patent document CN 106301384 A "a signal reconstruction method based on block compression sensing": the original signal is evenly divided into sub-blocks, and the sub-blocks are sparsely transformed and filtered; the obtained sub-signals are observed again to obtain an observation vector. The sub-signal is recovered by using the observation vector and the measurement matrix, and the reconstructed signal is obtained by linearly combining the sub-signals. Block-compressed sensing has the advantage that the storage measurement matrix is small and the reconstruction algorithm is simpler to calculate; however, the algorithm reconstructs each sub-signal separately and then linearly combines, which is not robust and easily causes reconstructed image block effects.

In the literature "Compressive sampling based on frequency saliency for remote sensing imaging", in order to improve the effect of block compression sensing reconstruction, the measurement rate is allocated according to the saliency information of the block, which effectively improves the reconstruction effect. However, the algorithm uses the OMP algorithm in reconstruction, and does not make full use of the prior knowledge of the image, and the set sparsity value has a great influence on the reconstruction effect.

At present, the block-compression sensing algorithm adopts uniform block, which reduces the computational complexity, but also destroys the original structure between image blocks, resulting in block-blocking perceptual sensing algorithm. Therefore, many scholars have proposed A lot of filtering methods smooth the difference between the block and the block, and then remove the block effect, but this also causes the image restoration structure to be destroyed while the block effect is removed, and the image reconstruction effect is reduced.

The non-local low-rank regularization compressed sensing reconstruction algorithm proposed in the literature "Compressive Sensing via Nonlocal Low-Rank Regularization" makes a very good reconstruction effect by fully utilizing the non-local similarity of the image, but the algorithm is The compression end uses an overall uniform measurement and does not make full use of the measurement information. In order to achieve a better reconstruction effect, the present invention performs reconstruction using the non-local low rank method at the decoding end.

Summary of the invention

In order to solve the above-mentioned defects existing in the prior art, to solve the problem of block compressed sensing reconstruction and to improve the peak signal to noise ratio of the reconstructed image, an object of the present invention is to provide a compressed sensing method based on gradient block adaptive measurement.

The technical solution adopted by the present invention to solve the technical problem is: in order to reduce the contradiction between the deblocking effect and the retained restoration structure of the blocking method, combined with the image geometric structure, an uneven blocking strategy is provided, and the region with similar structure is exhausted. Possible division into a block reduces the overall number of blocks of the image and the generation of blockiness, thereby reducing the structural loss in the deblocking effect. In the measurement, the adaptive measurement is performed according to the smoothness of the uneven block. When the overall measurement rate is constant, the measurement rate of the smooth region is reduced according to the smoothness adaptively, and the measurement rate of the complex region is increased to improve the effective information. Utilization.

Including the following steps:

(1) inputting a to-be-processed image, wherein the digitized matrix of the image to be processed is X∈R ^m×n , where m and n are the number of rows and columns of the image matrix, respectively;

(2) The image is divided into blocks, and the image gradient is divided into d uneven block images, and all the image blocks are expanded into a column vector x _j , where j={1, 2, ..., d}, combined into one Column vector x=[x ₁ ;x ₂ ;...;x _d ];

(3) assigning a measurement rate M _{j of} each block according to the sampling rate M required for the image and the smoothness of each block, j={1, 2, ..., d};

(4) Divide the blocks with the same number of elements into one class, and generate a Gaussian random orthogonal matrix of the same size. Combine the measurement rate M _{j of} each block and select the front round of the Gaussian random orthogonal matrix (M _j ×length(x _j )) as the measurement matrix φ _j ; measure the data to obtain the measured value y = Φx = diag (φ ₁ , φ ₂ , ..., φ _d ) [x ₁ ; x ₂ ;...; x _d ] ;

(5) Reconstruct the original signal by using the measured value, block measurement rate, line segmentation position and measurement matrix, and solve the problem by using the non-local low rank regularized compressed sensing reconstruction algorithm to output the reconstructed image;

x _j represents a column vector in which the block image matrix after the block is expanded into columns, and j is the order of the image blocks, and is a total of d blocks;

Φ _j represents a measurement matrix of x _j , and the number of rows is selected according to the sampling rate, and the number of columns is determined by the size of the image block;

y _j is a measured value of a column vector for each image block;

The blocking algorithm in the step (2) is as follows:

Step 1: Calculate the image X∈R ^m×n vertical direction gradient matrix D _v and horizontal direction gradient matrix D _h :

Where X(i,j) represents the pixel value of i row j column in image matrix X, and D _v (i,j), D _v (i,j) respectively represent the vertical of i row j column in image matrix X Directional gradient and horizontal gradient.

Step 2: The image X is divided into basic blocks of r × r which do not overlap each other, and r is a multiple of 2, and if it is less than r × r, it is complemented by 0.

Step 3: Perform a pooling operation on the vertical direction gradient matrix D _v and the horizontal direction gradient matrix D _h as a unit, respectively, to obtain a vertical direction smoothing matrix And horizontal direction smoothing matrix The vertical and horizontal smoothness of each basic block is calculated according to the gradient in the vertical direction and the gradient in the horizontal direction, respectively:

Where S _v (p, q), S _h (p, q) respectively represent the vertical smoothness and horizontal smoothness of the p-th row q-th column basic block in units of the basic block; ε is a constant, avoiding the denominator Zero, when taking ε = 1, S _v (p, q), S _h (p, q) ∈ (1, 255), the larger the value, the better the smoothness.

Step 4: Divide the elements of each row of S _v into two categories, set a threshold T _v , S _v (i, j) < T _{v is a} non-smooth block, and S _v (i, j) ≥ T _v is a smooth block, N _v (i) is the number of smoothing blocks in the ith row, and the row segmentation parameters are calculated:

Indicates the proportion of the smoothing block of the i-th row. Since the image structure has a regionality, the proportion of the number of smoothing blocks is used as a reference value for the row division.

Step 5: Perform row segmentation on the base block image from top to bottom according to the vertical smoothness matrix S _v , and R _v store the row index of the last row of the basic block in the row block, and length(R _v ) is the final row block number. .

5.1: Initial i=1, j=1, R _v (j)=1, σ _v is a constant, indicating the line block partitioning threshold.

5.2: i=i+1, judge |L _v (i)-L _v (i-1)|<σ _v , if it is true, indicating the smoothness of the vertical direction of the ith base block row and the i-1 base block row Within the threshold range, the i-th base block row is classified into the same row block containing the i-1th base block row, at which time R _v (j)=i; otherwise, the vertical smoothness of the two basic block rows is indicated. If it is not within the threshold, the i-th base block row is divided into the next row of blocks, at this time j=j+1, R _v (j)=i.

5.3: Judging If it is established, it indicates that the row block division is completed, and step 6 is performed. If not, execute 5.2.

Step 6: For each row block that is divided, the column segmentation is performed from left to right according to the horizontal smoothness matrix S _h , and the segmentation method is similar to the row segmentation method.

6.1: initial j=0, R _v (0)=0;

6.2: j = j + 1, S _h extracting j-th line block _{s h, s h = S h} (R v (j-1) +1: R v (j), :), each pair of s _h The elements of a column are divided into two categories, the threshold T _{h is set} , s _h (i, j) < T _{h is a} non-smooth block, and s _h (i, j) ≥ T _h is a smooth block, and N _h (k) represents s _h The kth column smoothes the number of blocks and calculates the column segmentation parameters:

Where L _h (k) represents the proportion of the smooth block of the kth column of s _h , and R _v stores the row index of the base block of the last row of the split line block,

_{R v (j) -R v (} j-1) s _h represents the number of blocks in each column base, which is equal to the number of elements in each column s _h,.

6.3: For the s _h column partition, R _h (j, k) stores the last column index of the kth column block dividing the jth row block. Length(R _h (j,:)) is the number of s _h column blocks. Since the smoothness of each row block is different, the number of divided column blocks of each row block is different.

6.4: Initial i=1, k=1, R _h (j, k)=1, σ _h is a constant, and the column block threshold is divided.

6.5: i=i+1, judge |L _h (k)-L _h (k-1)|<σ _h , if it is true, represent the kth basic block column and the k-1th base in the jth row block If the smoothness of the horizontal direction of the block column is within the threshold range, the kth base block column is classified into the same column block containing the k-1th basic block column, and then R _h (j, k)=i; otherwise, If the horizontal smoothness of the two basic block columns is not within the threshold, the kth basic block column is divided into the next column block, where k=k+1, R _h (j, k)=i.

6.6: Judgment If not, it indicates that the column division of the jth row block is not completed, and 6.5 is performed; if it is, judge j=length(R _v ), if not, execute step 6.2. If it is established, the whole partial block is completed, and R _h (j, k).

The measurement rate in the step (3) is an adaptive allocation algorithm, and the algorithm is as follows:

Step 1: Enter the overall measurement rate M of the image, and calculate the smoothness of all the basic blocks in the image block x _j , j=1,...,q:

Step 2: Assign measurement rate based on sub-block smoothness:

Where M ₀ >0 is the lowest measurement rate of the block.

The step (4) is performed by using a non-local low rank regularized compressed sensing reconstruction algorithm, and the column vector is reconstructed into an image and processed after reconstruction;

Since the above technical solution is adopted, the beneficial effects of the present invention are: by using the block compressed sensing image reconstruction algorithm of the present invention, the block effect caused by the block compression sensing is effectively reduced; and the existing adaptive The block measurement method has a higher peak signal-to-noise ratio and better image visual effects.

DRAWINGS

1 is a general flow chart of a method for compressive sensing based on gradient block adaptive measurement according to the present invention;

2 is a flow chart of an algorithm based on gradient uneven block of the present invention;

Detailed ways

A method for compressive sensing based on gradient block adaptive measurement according to the present invention will be described in detail below with reference to the accompanying drawings and a typical embodiment. The algorithm specifically includes the following parts:

The compression side compresses the image by using a gradient-based non-uniform block adaptive measurement method. The steps are as follows:

Input the image to be processed X∈R ^m×n , calculate the horizontal and vertical gradients of the image; pool the image gradient according to the basic block size to obtain the smoothness of the horizontal and vertical directions of the image basic block; according to the vertical direction The smoothness of the image is divided into line blocks from top to bottom, and then each line block is divided into column blocks from left to right according to the horizontal smoothness, thereby completing uneven block, and the r×r basic block is used in the present invention. Increase the split base unit and reduce the amount of calculation. Specific steps are as follows:

Step 1: Calculate the image X∈R ^m×n vertical direction gradient matrix D _v and horizontal direction gradient matrix D _h :

Where X(i,j) represents the pixel value of i row j column in image matrix X, and D _v (i,j), D _v (i,j) respectively represent the vertical of i row j column in image matrix X Directional gradient and horizontal gradient.

Step 2: The image X is divided into basic blocks of r × r which do not overlap each other, and r is a multiple of 2, and if it is less than r × r, it is complemented by 0.

Step 3: Perform a pooling operation on the vertical direction gradient matrix D _v and the horizontal direction gradient matrix D _h as a unit, respectively, to obtain a vertical direction smoothing matrix And horizontal direction smoothing matrix The vertical and horizontal smoothness of each basic block is calculated according to the gradient in the vertical direction and the gradient in the horizontal direction, respectively:

Where S _v (p, q), S _h (p, q) respectively represent the vertical smoothness and horizontal smoothness of the p-th row q-th column basic block in units of the basic block; ε is a constant, avoiding the denominator Zero, when taking ε = 1, S _v (p, q), S _h (p, q) ∈ (1, 255), the larger the value, the better the smoothness.

Step 4: Divide the elements of each row of S _v into two categories, set a threshold T _v , S _v (i, j) < T _{v is a} non-smooth block, and S _v (i, j) ≥ T _v is a smooth block, N _v (i) is the number of smoothing blocks in the ith row, and the row segmentation parameters are calculated:

Indicates the proportion of the smoothing block of the i-th row. Since the image structure has a regionality, the proportion of the number of smoothing blocks is used as a reference value for the row division.

Step 5: Perform row segmentation on the base block image from top to bottom according to the vertical smoothness matrix S _v , and R _v store the row index of the last row of the basic block in the row block, and length(R _v ) is the final row block number. .

5.1: Initial i=1, j=1, R _v (j)=1, σ _v is a constant, indicating the line block partitioning threshold.

5.2: i=i+1, judge |L _v (i)-L _v (i-1)|<σ _v , if it is true, indicating the smoothness of the vertical direction of the ith base block row and the i-1 base block row Within the threshold range, the i-th base block row is classified into the same row block containing the i-1th base block row, at which time R _v (j)=i; otherwise, the vertical smoothness of the two basic block rows is indicated. If it is not within the threshold, the i-th base block row is divided into the next row of blocks, at this time j=j+1, R _v (j)=i.

5.3: Judging If it is established, it indicates that the row block division is completed, and step 6 is performed. If not, execute 5.2.

Step 6: For each row block that is divided, the column segmentation is performed from left to right according to the horizontal smoothness matrix S _h , and the segmentation method is similar to the row segmentation method.

6.1: initial j=0, R _v (0)=0;

6.2: j = j + 1, S _h extracting j-th line block _{s h, s h = S h} (R v (j-1) +1: R v (j), :), each pair of s _h The elements of a column are divided into two categories, the threshold T _{h is set} , s _h (i, j) < T _{h is a} non-smooth block, and s _h (i, j) ≥ T _h is a smooth block, and N _h (k) represents s _h The kth column smoothes the number of blocks and calculates the column segmentation parameters:

Where L _h (k) represents the proportion of the smooth block of the kth column of s _h , and R _v stores the row index of the base block of the last row of the split line block,

_{R v (j) -R v (} j-1) s _h represents the number of blocks in each column base, which is equal to the number of elements in each column s _h,.

6.3: For the s _h column partition, R _h (j, k) stores the last column index of the kth column block dividing the jth row block. Length(R _h (j,:)) is the number of s _h column blocks. Since the smoothness of each row block is different, the number of divided column blocks of each row block is different.

6.4: Initial i=1, k=1, R _h (j, k)=1, σ _h is a constant, and the column block threshold is divided.

6.5: i=i+1, judge |L _h (k)-L _h (k-1)|<σ _h , if it is true, represent the kth basic block column and the k-1th base in the jth row block If the smoothness of the horizontal direction of the block column is within the threshold range, the kth base block column is classified into the same column block containing the k-1th basic block column, and then R _h (j, k)=i; otherwise, If the horizontal smoothness of the two basic block columns is not within the threshold, the kth basic block column is divided into the next column block, where k=k+1, R _h (j, k)=i.

6.6: Judgment If not, it indicates that the column division of the jth row block is not completed, and 6.5 is performed; if it is, judge j=length(R _v ), if not, execute step 6.2. If it is established, the whole partial block is completed, and R _h (j, k).

Then, according to the sampling rate M required for the compressed image and the smoothness of each block, the measurement rate M _j , j={1, 2, . . . , d} of each block is allocated, and the specific steps are as follows:

Step 1: Enter the overall measurement rate M of the image, and calculate the smoothness of all the basic blocks in the image block x _j , j=1,...,q:

Step 2: Assign measurement rate based on sub-block smoothness:

Where M ₀ >0 is the lowest measurement rate of the block.

Finally, the blocks with the same number of elements are divided into one class, and a Gaussian random orthogonal matrix of the same size is generated. Combining the measurement rate M _{j of} each block, the front round (M _j ×length(x _j )) of the Gaussian random orthogonal matrix is selected. The line is used as the measurement matrix φ _j ; the data is measured to obtain a measured value y = Φx = diag (φ ₁ , φ ₂ , ..., φ _d ) [x ₁ ; x ₂ ;...; x _d ];

The decompression end adopts a non-local low rank regularization compressed sensing reconstruction algorithm. The specific steps are as follows:

Step 1: According to the measured value, block measurement rate, line segmentation position and measurement matrix, the initial estimated image x ⁽⁰⁾ is obtained by DCT soft threshold algorithm;

Step 2: Initialization, k=1, x ^(k) = x ⁽⁰⁾ , k represents the number of iterations. The x _{1 is} divided into image reference blocks of size 6 × 6 in steps of 5, and 44 most similar similar blocks are searched in the image x ₁ according to each reference block. The criterion for judging the similarity, the present invention uses the cosine angle function of the vector to judge:

among them, a column vector representing the i-th reference block of the image x ^(k) , Represents a column vector of an image block of size 6 × 6 in image x ^(k) , and c represents an image block Reference block The similarity, the larger the value, the greater the similarity, the maximum value is 1. The 45 similar block column vectors including the reference block are composed into a similar block matrix X _i , where X _i =R _i x=[R _i0 x,R _i1 x,...,R _iq x], and R represents extraction Block operation, q=45.

Step 3: Create a model. The low rank model of the similar block matrix X _i is:

among them, Is the variance of Gaussian noise.

The regular term for the low rank model as the reconstruction model is:

Where η and λ are regularization parameters, and are set as constants during reconstruction; L(L _i , ε) represents a rank substitution function, and a kernel function is used as a rank substitution function.

Using the alternating solution method, convert to:

Step 4: Solution similar blocks low-rank matrix L _i X _i of the matrix, Singular Value shrinkage algorithm available:

among them, σ _j indicates The jth singular value, UΣV ^T is The singular value decomposition matrix, (x) ₊ = max{x, 0}.

Step 5, solving the reconstructed image The termination condition is judged if the output reconstructed image is satisfied. Otherwise, when mod(k, T)=0, the most similar 45 similar blocks are re-searched for each reference block to form a similar block matrix X _i , and the process _proceeds to step 3. among them, Representing an operation of restoring a similar block in the i-th similar block group to the original image position,

It is to be understood that the foregoing description is only a particular embodiment of the present invention, and the invention is not limited to the specific structures shown and described herein.

Claims (4)

A compressed sensing method based on gradient block adaptive measurement, comprising the steps of: Including the following steps: (1) inputting a to-be-processed image, wherein the digitized matrix of the image to be processed is X∈R ^m×n , where m and n are the number of rows and columns of the image matrix, respectively; (2) Blocking the image, dividing into d uneven image blocks based on the image gradient, and expanding all the image blocks into a column vector x _j , where j={1, 2, ..., d}, combined into A column vector x=[x ₁ ;x ₂ ;...;x _d ]; (3) assigning a measurement rate M _{j of} each block according to the sampling rate M required for the compressed image and the smoothness of each block, j = {1, 2, ..., d}; (4) Divide the blocks with the same number of elements into one class, and generate a Gaussian random orthogonal matrix of the same size. Combine the measurement rate M _{j of} each block and select the front round of the Gaussian random orthogonal matrix (M _j ×length(x _j )) as the measurement matrix φ _j ; measure the data to obtain the measured value y = Φx = diag (φ ₁ , φ ₂ , ..., φ _d ) [x ₁ ; x ₂ ;...; x _d ] ; (5) Reconstruct the original signal by using the measured value, block measurement rate, line segmentation position and measurement matrix, and solve the problem by using the non-local low rank regularized compressed sensing reconstruction algorithm to output the reconstructed image. The method according to claim 1, wherein the blocking algorithm in the step (2) is as follows: Step 1: Calculate the image X∈R ^m×n vertical direction gradient matrix D _v and horizontal direction gradient matrix D _h : Where X(i,j) represents the pixel value of i row j column in image matrix X, and D _v (i,j), D _v (i,j) respectively represent the vertical of i row j column in image matrix X Directional gradient and horizontal gradient; Step 2: Divide the image X into basic blocks of r×r that do not overlap each other, and take a multiple of 2, if less than r×r, fill with 0; Step 3: Perform a pooling operation on the vertical direction gradient matrix D _v and the horizontal direction gradient matrix D _h as a unit, respectively, to obtain a vertical direction smoothing matrix And horizontal direction smoothing matrix The vertical and horizontal smoothness of each basic block is calculated according to the gradient in the vertical direction and the gradient in the horizontal direction, respectively: Where S _v (p, q), S _h (p, q) respectively represent the vertical smoothness and horizontal smoothness of the p-th row q-th column basic block in units of the basic block; ε is a constant, avoiding the denominator Zero, when taking ε=1, S _v (p,q),S _h (p,q)∈(1,255], the larger the value, the better the smoothness; Step 4: Divide the elements of each row of S _v into two categories, set a threshold T _v , S _v (i, j) < T _{v is a} non-smooth block, and S _v (i, j) ≥ T _v is a smooth block, N _v (i) is the number of smoothing blocks in the ith row, and the row segmentation parameters are calculated: Indicates the proportion of the smoothing block in the i-th row. Because the image structure has regionality, the proportion of the number of smoothing blocks is used as the reference value of the row segmentation; Step 5: Perform row segmentation on the base block image from top to bottom according to the vertical smoothness matrix S _v , and R _v store the row index of the last row of the basic block in the row block, and length(R _v ) is the final row block number. ; 5.1: initial i=1, j=1, R _v (j)=1, σ _v is a constant, indicating a row block segmentation threshold; 5.2: i=i+1, judge |L _v (i)-L _v (i-1)|<σ _v , if it is true, indicating the smoothness of the vertical direction of the ith base block row and the i-1 base block row Within the threshold range, the i-th base block row is classified into the same row block containing the i-1th base block row, at which time R _v (j)=i; otherwise, the vertical smoothness of the two basic block rows is indicated. If it is not within the threshold, the i-th base block row is divided into the next row of blocks, at this time j=j+1, R _v (j)=i; 5.3: Judging If it is established, it indicates that the row block division is completed, and step 6 is performed. If not, execute 5.2; Step 6: Perform column segmentation from left to right according to the horizontal smoothness matrix S _h for each of the divided block blocks, and the segmentation method is similar to the row segmentation method; 6.1: initial j=0, R _v (0)=0; 6.2: j = j + 1, S _h extracting j-th line block _{s h, s h = S h} (R v (j-1) +1: R v (j), :), each pair of s _h The elements of a column are divided into two categories, the threshold T _{h is set} , s _h (i, j) < T _{h is a} non-smooth block, and s _h (i, j) ≥ T _h is a smooth block, and N _h (k) represents s _h The kth column smoothes the number of blocks and calculates the column segmentation parameters: Where L _h (k) represents the proportion of the k-th column smoothing block of s _h , R _v stores the row index of the last row of the basic block of the segmentation row block, and R _v (j)-R _v (j-1) represents s _h The number of basic blocks in each column, the value of which is equal to the number of elements in each column of s _h ; 6.3: For the s _h column partition, R _h (j, k) stores the last column index of the kth column block dividing the jth row block. Length(R _h (j,:)) is the number of s _h column blocks. Since the smoothness of each row block is different, the number of divided column blocks of each row block is different; 6.4: initial i=1, k=1, R _h (j, k)=1, σ _h is a constant, and the column block threshold is divided; 6.5: i=i+1, judge |L _h (k)-L _h (k-1)|<σ _h , if it is true, represent the kth basic block column and the k-1th base in the jth row block If the smoothness of the horizontal direction of the block column is within the threshold range, the kth base block column is classified into the same column block containing the k-1th basic block column, and then R _h (j, k)=i; otherwise, If the horizontal smoothness of the two basic block columns is not within the threshold, the kth basic block column is divided into the next column block, where k=k+1, R _h (j, k)=i; 6.6: Judgment If not, it indicates that the column division of the jth row block is not completed, and 6.5 is performed; if it is, judge j=length(R _v ), if not, execute step 6.2. If it is established, the whole partial block is completed, and R _h (j, k). The method according to claim 1, wherein the measurement rate in the step (3) adopts an adaptive allocation algorithm, and the algorithm is as follows: Step 1: Enter the overall measurement rate M of the image, and calculate the smoothness of all the basic blocks in the image block x _j , j=1,...,q: Step 2: Assign measurement rate based on sub-block smoothness: Where M ₀ >0 is the lowest measurement rate of the block. The gradient sensing adaptive measurement-based compressed sensing method according to claim 1, wherein the step (5) adopts a similar block judgment criterion in a non-local low rank regularized compressed sensing reconstruction algorithm. : among them, a column vector representing the i-th reference block of the image x ^(k) , Represents a column vector of an image block of size 6 × 6 in image x ^(k) , and c represents an image block Reference block The similarity, the larger the value, the greater the similarity, the maximum value is 1.