An adaptive block-based compressed sensing image reconstruction method based on texture information

Abstract

The present invention discloses an adaptive block-based compressed sensing image reconstruction method based on texture information, belonging to the field of image processing technology. The method comprises the following steps: dividing the original image into blocks; extracting texture information from each image block and calculating its texture information weight, wherein the texture information includes grayscale entropy, standard deviation, and edge information; clustering each image block according to the difference in texture information weight; calculating the adaptive sampling rate of each type of image block, and constructing a measurement matrix for each type of image block based on the adaptive sampling rate for measurement; and reconstructing the image using a smoothed projection Landweber reconstruction algorithm based on a superposition structure block method to obtain large-size reconstructed image blocks for splicing. Compared with the prior art, the present invention effectively suppresses block artifacts and improves the subjective quality of image reconstruction and peak signal-to-noise ratio while maintaining the total sampling rate unchanged. It also achieves a reasonable adaptive allocation of sampling rates and has excellent practicality.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to an adaptive block-based compressed sensing image reconstruction method based on texture information. Background Art

[0002] Since its introduction, compressed sensing theory has become a hot topic in information research, attracting the attention of numerous scholars both domestically and internationally. Traditional sampling theorems require that the sampling frequency be at least twice the signal bandwidth to fully recover the original signal. In compressed sensing, information sampling and compression are performed simultaneously, reducing storage space and computational complexity. Furthermore, this theory overcomes the limitations of traditional sampling theorems by lowering the sampling rate and enabling rapid signal reconstruction with fewer samples.

[0003] When traditional compressed sensing theory is applied to two-dimensional images, the image sampling process needs to obtain the entire image at one time. Although the sampling rate has been reduced to a certain extent compared with the Nyquist sampling rate, the scale of the observation matrix for the entire image observation is still huge, generally around 10 4 ~10 6 The magnitude of the data is large, which makes storage and calculation inconvenient and limits the real-time performance of image perception.

[0004] In 2006, Gan L proposed a block compressed sensing (BCS) method for images. BCS uses the same block compressible sampling operator to acquire images, which reduces the storage required for the compressible sampling operator during the compressed sensing process, making image block reconstruction easier and significantly reducing storage space and computation time. Compared to directly processing the entire image, BCS can reduce the size of the observation matrix and achieve rapid image reconstruction. However, using the same observation matrix for each image block ignores the differences in information distribution between blocks, which affects reconstruction performance and makes blocking artifacts more likely to occur. Furthermore, since the same sampling rate is used for image blocks with different texture information richness, blocks with less texture information are oversampled and blocks with rich texture information are undersampled, limiting the quality of the reconstructed image. Summary of the Invention

[0005] In response to the above-mentioned problems, the present invention provides an adaptive block compressed sensing image reconstruction method based on texture information. Without increasing the total sampling rate at the measurement end and the algorithm complexity at the reconstruction end, image blocks are clustered according to the texture information of the image and a secondary sampling rate is allocated to achieve adaptive sampling. By proposing a superimposed block structure and performing reconstruction at the reconstruction end, the subjective quality and peak signal-to-noise ratio of the reconstructed image can be improved, and the blocking effect can be reduced.

[0006] The technical solutions adopted by the present invention to solve the above technical problems are as follows:

[0007] An adaptive block-based compressed sensing image reconstruction method based on texture information specifically includes the following steps:

[0008] Step 1: Divide the original image into blocks;

[0009] Step 2: extracting texture information of each image block and calculating its texture information weight, wherein the texture information of each image block includes grayscale entropy, standard deviation, and edge information;

[0010] Step 3: clustering the image blocks according to the texture information weight differences of the image blocks calculated in step 2;

[0011] Step 4: Calculate the adaptive sampling rate of each type of image block, and construct a measurement matrix of each type of image block according to the adaptive sampling rate to perform measurement;

[0012] Step 5: Reconstruct the image using the smooth projection Landweber reconstruction algorithm based on the superposition structure block method to obtain large-size reconstructed image blocks for splicing.

[0013] Furthermore, the specific content of dividing the original image into blocks in step 1 includes: dividing the image of size N=W×H into m image blocks of the same size B×B and non-overlapping, and the column vector form of the j-th image block is x j , j=1,2,...,m,m=N / B 2 .

[0014] Furthermore, the content of calculating the texture information weight of each image block in the step 2 includes: after dividing the image into blocks, calculating the grayscale entropy h of the image block j according to the empirical formula j ,Std j , and edge information E j ; Then calculate the grayscale entropy, standard deviation and weight of edge information of each image block to the corresponding texture information of all image blocks, and get the grayscale entropy weight Standard deviation weight and edge weights The above texture information reflects the richness of texture information of each image block, and is used as the basis for subsequent image block clustering and calculation of adaptive sampling rate.

[0015] Furthermore, the specific process of clustering each image block in step 3 is as follows: using the texture information weight of each image block calculated in step 2 as the input of the K-means++ clustering algorithm, clustering each image block according to the richness of texture information using the K-means++ algorithm, and classifying each image block from less to more rich in texture information into:

[0016]

[0017] Furthermore, the specific contents of calculating the adaptive sampling rates of various image blocks in step 4 include:

[0018] Step 41): Set the total sampling rate SR∈[0.1,0.5] of the image;

[0019] Step 42): Calculate the pre-sampling rate SR_Pre of each image block, the pre-sampling rate SR_Pre of image block j j Defined as:

[0020]

[0021] in are the texture information weights of each image block; W and H are the width and height of the original image respectively; B is the size of the image after segmentation; SR is the total image sampling rate.

[0022] Step 43): According to the pre-sampling rate SR_Pre of each image block and the clustering result obtained in step 3, the pre-sampling of the same type of image blocks is first summed and then averaged to obtain the adaptive sampling rate SR of each type of image block. j ;

[0023] Step 44): Adaptive sampling rate SR for each type of image block j Perform secondary distribution

[0024] Set the sampling rate upper limit SR max , traverse each image block, if the sampling rate of an image block exceeds the upper bound of the sampling rate, calculate the part that exceeds the upper bound of the sampling rate:

[0025] SR plus =SR plus +(SR j -SR max )

[0026] Among them SR max is the upper bound of the sampling rate; then SR plus Evenly distribute to all image blocks. If there are still image blocks with sampling rates exceeding the upper limit of the sampling rate, repeat this step until all image blocks have SR j ≤SR max .

[0027] Furthermore, the upper limit of the sampling rate is set to 0.8.

[0028] Furthermore, the specific content of constructing the measurement matrix of various image blocks is: using a size of B 2 ×B 2The orthogonal Gaussian random matrix is ​​used as the measurement matrix Φ; according to the adaptive sampling rate of the flat block, transition block and texture block, the number of measurements of each block is determined and their respective measurement matrices are generated. and Take measurements.

[0029] Furthermore, the specific process of the smooth projection Landweber reconstruction algorithm based on the superposition structure block method described in step 5 is as follows:

[0030] Assume that the block sizes during measurement and reconstruction are B1 and B2 respectively, and to achieve the superposition structure,

[0031] B1 and B2 satisfy the following relationship:

[0032] B1<B2 and B2=2 l ×B1 l=1,2,3,...,L

[0033] Where l is a positive integer. The superposition block structure first divides the two-dimensional image X into m non-overlapping blocks of size N during measurement. B =B1×B1 image block X bj (j=1,2,...,m), and transform it into a B ×1 one-dimensional vector x bj ;

[0034] Then by constructing a B ×N B Measurement matrix Φ bj , where M B =SR j ×B is the total number of samples of each image block; the size of each block is M B ×1 measurement vector y bj , which is expressed as follows:

[0035] y bj =Φ bj ·x bj j=1,2,...,m

[0036] At the reconstruction end, the measurement value vector y at the measurement end bj , the measurement matrix Φ bj and the one-dimensional vector x bj Splicing and reorganizing into large-size blocks; the reconstruction end sets l = 1, and the splicing process is as follows:

[0037]

[0038] where Φ ∧ are four measurement matrices Φ bj The diagonal matrix formed by the concatenation of y and Φ ∧The image is reconstructed as the input of the smooth projection Landweber reconstruction algorithm to obtain a large-size reconstructed image block.

[0039] Furthermore, it also includes judging whether the large-size reconstructed image block obtained in step five meets the peak signal-to-noise ratio requirement for the reconstructed image. If it meets the requirement, the image reconstruction process is terminated to obtain the reconstructed image, and all large-size reconstructed image blocks are spliced ​​to obtain the entire reconstructed image.

[0040] The technical solution of the present invention can produce the following technical effects:

[0041] The adaptive block-based compressed sensing image reconstruction method based on texture information proposed in this paper clusters image blocks based on their texture information and allocates a secondary sampling rate, without increasing the total sampling rate at the measurement end or the algorithm complexity at the reconstruction end. This achieves reasonable sampling rate adaptation. By proposing a superimposed block structure and performing reconstruction at the reconstruction end, this adaptive block-based compressed sensing image reconstruction method has been experimentally verified to suppress blocking artifacts in the reconstructed image and effectively improve the subjective reconstruction quality and peak signal-to-noise ratio of the reconstructed image while maintaining the same total sampling rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is an overall flow chart for implementing the adaptive block-based compressed sensing image reconstruction method based on texture information of the present invention;

[0043] Figures 2a-2e They are the Lenna, Peppers, Mandrill, Goldhill, and Barbara image sets tested in the embodiments of the present invention respectively;

[0044] Figure 3 This is a comparison diagram of Barbara image block clustering at the measurement end according to an embodiment of the present invention;

[0045] Figure 4 is a flow chart of a method for reconstructing a superimposed structure according to the present invention;

[0046] FIG5 is a diagram showing the reconstruction results of the Barbara image by the present invention and the existing BCS-SPL method and a partially enlarged diagram; Figure 5a From left to right are the overall comparison diagrams of the original image of Barbara, the image of Barbara using the BCS-SPL algorithm, and the image of Barbara reconstructed using the present invention; Figure 5b From left to right are the original image of Barbara, the partially enlarged image of Barbara image using the BCS-SPL algorithm, and the reconstructed image of Barbara image using the present invention;

[0047] FIG6 is a diagram showing the reconstruction results of Lenna images by the present invention and the existing BCS-SPL method and a partially enlarged diagram; Figure 6a From left to right are the overall comparison diagrams of the original Lenna image, the Lenna image after BCS-SPL algorithm, and the Lenna image reconstructed by the present invention; Figure 6b From left to right are the original Lenna image, the Lenna image obtained by the BCS-SPL algorithm, and the partially enlarged images of the Lenna image reconstructed by the present invention;

[0048] FIG7 is a diagram showing the effect of the present invention on the deblocking effect of the Goldhill image; Figure 7a Comparison of the reconstructed images of ABCS-SPL and the present invention when the total sampling rate is 0.3; Figure 7b Comparison of the reconstructed images of ABCS-SPL and the present invention when the total sampling rate is 0.5;

[0049] Figure 8 This is a line graph comparing the peak signal-to-noise ratio of images reconstructed by the present invention and the existing algorithm. DETAILED DESCRIPTION

[0050] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them.

[0051] like Figure 1 As shown, the overall process of implementing the adaptive block-based compressed sensing image reconstruction method based on texture information described in the present invention is as follows:

[0052] Step 1: Divide the test image set into several image blocks of equal size and without overlapping. The specific process is as follows:

[0053] This embodiment Figures 2a-2e , five standard test set images of Lenna, Peppers, Mandrill, Goldhill, and Barbara, all of which are 512×512 in size, are used as original images; the block size of this embodiment is 32×32 but not limited to this value, and the number of blocks is 256; the column vector form of the j-th image block is x j , j=1,2,...,256.

[0054] Step 2: Calculate the texture information weights of the 256 image blocks in the test image set. The specific process is as follows:

[0055] 21) Calculate the grayscale entropy h of image block j according to the empirical formula j ,Std j , and edge information Ej ; Among them, the edge information E j is the number of edge points of each image block calculated using the Sobel edge detection method with the Scharr convolution kernel. The threshold of the Sobel edge detection in this embodiment is set to 450 based on experience;

[0056] 22) Calculate the grayscale entropy weight of each image block separately Standard deviation weight and edge weights And there is

[0057] Step 3: clustering the image blocks according to the texture information weight differences of the image blocks calculated above;

[0058] The texture information weights calculated in step 2 are used as input to the K-means++ clustering algorithm. The K-means++ algorithm is used to cluster the image blocks according to the richness of texture information. The image blocks are classified from less to more rich in texture information into the following categories:

[0059]

[0060] The clustering results are detailed in Simulation 2 below.

[0061] Step 4: Calculate the adaptive sampling rate of each type of image block, and construct the measurement matrix of each type of image block according to the adaptive sampling rate for measurement. The specific process is as follows:

[0062] 41) In this embodiment, the total sampling rate of the image is set to SR∈[0.1,0.5];

[0063] 42) Calculate the pre-sampling rate SR_Pre of each image block, the pre-sampling rate SR_Pre of image block j j Defined as:

[0064]

[0065] in are the texture weights of each image block; W and H are the width and height of the original image, respectively; B is the size of the image after segmentation. SR is the total image sampling rate. W = H = 512, B = 32.

[0066] 43) According to the clustering results, calculate the adaptive sampling rates of flat blocks, transition blocks, and texture blocks respectively. Let m smooth ,m transition ,m texture Represents the number of flat blocks, transition blocks, and texture blocks respectively. The adaptive sampling rates of the three types of image blocks are:

[0067]

[0068]

[0069]

[0070] 44) Set the upper limit of the sampling rate SR max , traverse each image block, if there is an image block sampling rate SR j If the sampling rate exceeds the upper limit, the SR of the part exceeding the upper limit of the sampling rate is calculated. plus =SR plus +(SR j -SR max ), where SR max The upper bound of the sampling rate is set to 0.8; then SR plus Evenly distribute to all image blocks. If there are still image blocks with sampling rates exceeding the upper limit of the sampling rate, repeat this step 44) until all image blocks have SR j ≤SR max This achieves a secondary distribution of the sampling rate and obtains the adaptive sampling rate of each type of image block.

[0071] 45) Construct the measurement matrix of each image block according to the adaptive sampling rate of each image block. 2 ×B 2 The orthogonal Gaussian random matrix is ​​used as the measurement matrix Φ. The adaptive sampling rates of flat blocks, transition blocks and texture blocks are used to determine the number of measurements and generate their respective measurement matrices. and Take measurements;

[0072] Where B=32.

[0073] Step 5: Reconstruct the image using the smooth projection Landweber reconstruction algorithm based on the superposition structure block method to obtain large-size reconstructed image blocks for splicing; Figure 4 Flowchart of the method for reconstructing a superimposed structure of the present invention, the specific process of the method is:

[0074] Assume that the dimensions during measurement and reconstruction are B1 and B2 respectively. To achieve the superposition structure, B1 and B2 satisfy the following relationship:

[0075] B1<B2 and B2=2 l ×B1 l=1,2,3,...,L

[0076] In this embodiment, B1=32, l=1, B2=64, the superimposed block structure first divides the two-dimensional image X into m non-overlapping blocks of size N during measurement. B =B1×B1 image block X bj(j=1,2,...,256), and convert it into a size of N B ×1 one-dimensional vector x bj ; Then construct a size of M B ×N B Measurement matrix Φ bj , where M B =SR j ×B is the total number of samples of each image block; to get the size of each block M B ×1 measurement value vector y bj . It is expressed as follows:

[0077] y bj =Φ bj ·x bj j=1,2,...,m

[0078] Among them, M B =M smooth orM B =M transition orM B =M texture At the reconstruction end, the measurement value vector y of the measurement end is bj , the measurement matrix Φ bj and the one-dimensional vector x bj Splice and reassemble into large-size blocks. Set l = 1, and the splicing process is as follows:

[0079]

[0080] where Φ ∧ are the four measurement matrices Φ bj The diagonal matrix formed by the concatenation of y and Φ ∧ The image is reconstructed as the input of the smooth projection Landweber reconstruction algorithm to obtain a large-size reconstructed image block; then, it is determined whether the obtained reconstructed image block meets the peak signal-to-noise ratio requirement for the reconstructed image. If it does, the image reconstruction process is terminated to obtain the reconstructed image; finally, all the large-size reconstructed image blocks are spliced ​​to obtain the entire reconstructed image.

[0081] The experimental simulation content of the adaptive block compressed sensing image reconstruction method based on texture information of the present invention is as follows:

[0082] The computer hardware configuration in the simulation experiment is Intel(R) Core(TM) i5-9300HF processor, 2.40GHz, 16.0GB memory, Windows 10 system laptop, and the programming language is Matlab.

[0083] Simulation 1: Clustering image blocks based on image texture information weights, Figure 3This example tests image clustering after dividing the Barbara image into 8×8 blocks. Black blocks represent flat blocks, gray blocks represent transition blocks, and white blocks represent texture blocks. Comparing the original image with the original image reveals that the clothing, pants, headscarf, and tablecloth areas of Barbara's image contain rich texture information, where texture blocks are concentrated. In contrast, the floor, walls, and table legs contain less texture information and are more concentrated with flat blocks. This comparison demonstrates that the image block clustering method presented in this paper is quite accurate.

[0084] Simulation 2: Reconstruction results and partial magnification comparison of the Barbara and Lenna images by the present invention and the existing classic block-based compressed sensing algorithm BCS-SPL method at a total sampling rate of 0.3. Figure 5a 、 Figure 5b From left to right in the figure are the original image of Barbara, the overall comparison image of the image reconstructed by the BCS-SPL algorithm and the present invention, and the local magnified comparison image. Figure 6a 、 Figure 6b From left to right in the middle are the original Lenna image, the overall comparison image and the local magnified comparison image of the reconstructed image using the BCS-SPL algorithm and the present invention. Figure 5b It can be seen that for stripe regions with more complex textures, the reconstructed image texture effect of the present invention is clearer. Figure 6b It can be seen that the image reconstruction effect of the present invention is closer to the original image, and the boundary between the pupil and the eye black is clearer. Compared with the enlarged image of the BCS-SPL algorithm, the texture area of ​​the present invention is clearer and more details and edges are preserved, and the reconstruction effect is closer to the original image.

[0085] Simulation 3: Deblocking Effect Comparison of the Present Invention

[0086] The adaptive sampling is combined with the BCS-SPL algorithm, and the improved algorithm is called ABCS-SPL. The image reconstructed by the present invention is compared with that by ABCS-SPL. Figure 7a and Figure 7b Comparison of the reconstructed images of ABCS-SPL and the present invention when the total sampling rates are 0.3 and 0.5 respectively: It can be seen that ABCS-SPL produces serious blocking effects regardless of the low sampling rate of 0.3 or the high sampling rate of 0.5, while the present invention effectively suppresses the blocking effect.

[0087] Simulation 4, Peak Signal-to-Noise Ratio (PSNR) comparison of reconstructed images

[0088] Figure 8Taking the Lenna image as an example, the peak signal-to-noise ratio (PSNR) of the reconstructed image of the existing algorithm and the present invention at a total sampling rate of 0.1-0.5 is compared in a broken line graph. Tables 1 to 4 respectively show the numerical comparison results of the peak signal-to-noise ratio (PSNR) of the reconstructed image of the Pepper, Barbaras, Goldhill, and Mandrill images using the existing algorithm and the present invention at a total sampling rate of 0.1-0.5. The existing algorithms are the classic block-based compressed sensing algorithm BCS-SPL, the multi-scale block-based compressed sensing algorithm MS-BCS-SPL, and the methods in reference [1], [2], and [3].

[0089] Table 1 Comparison of PSNR results of this method and existing algorithms for Peppers images

[0090]

[0091] Table 2 Comparison of PSNR results of this method and existing algorithms for Goldhill images

[0092]

[0093] Table 3 Comparison of PSNR results of this method and existing algorithms for Barbara image

[0094]

[0095] Table 4 Comparison of PSNR results between this method and existing algorithms for Mandrill images

[0096]

[0097]

[0098] Literature[1]:LI R,DUANX,LVY.Adaptive compressive sensingofimages usingerrorbetweenblocks[J].InternationJournal ofDistributed SensorNetworks,2018,14(6):1-8.

[0099] Literature [2]: Li R, GanZ L, ZhuXC. Afastcompressed-sensing image reconstruction algorithm based on best linear estimate. JElectron InformTechnol, 2012, 34: 3006–3012

[0100] Literature [3]: SHI Cuiping, WANG Liguo, NA Yujing, et al. Blockcompressivesensingmethodbasedonadaptive samplingand smoothprojection[J]. Journal of Harbin Engineering University, 2019, 41(6):877-883

[0101] Figure 8 It can be seen that at the total sampling rate SR = [0.1, 0.5], the image reconstruction effect of the present invention is better overall. Compared with BCS-SPL and MS-BCS-SPL, the PSNR is improved by an average of about 3.1-4.9dB at the low and medium sampling rates [0.1, 0.3], and by an average of about 4.9-7.8dB at the high sampling rates [0.4, 0.5]. Compared with the algorithms in references [1]-[3], the PSNR is improved by an average of about 2.2-6.1dB at the low and medium sampling rates [0.1, 0.3], and by an average of about 0.6-3.4dB at the high sampling rates [0.4, 0.5].

[0102] As can be seen from Tables 1 to 4, for Peppers images that contain textures and have fewer edges and are relatively flat overall, whether at low or high sampling rates, the increase in PSNR value for each 0.1 increase in sampling rate remains basically between 3.8 and 4.5 dB compared with other algorithms.

[0103] For images with rich texture information and edges, such as Barbara and Mandrill, the PSNR values ​​at low sampling rates SR = [0.1, 0.3] are improved by an average of 7.3-7.6dB and 6.4-8.8dB compared to other algorithms. This is because the number of samples is small for low sampling rates, and the present invention can allocate the sampling rate more reasonably and accurately, increasing the sampling rate of texture blocks at the expense of reducing the sampling rate of flat blocks, thereby improving the reconstruction quality of texture blocks. Since the proportion of texture blocks in these images is relatively large, the overall PSNR improvement value of the image is relatively high. At high sampling rates SR = [0.4, 0.5], the PSNR values ​​are improved by an average of 6dB and 5.3dB-5.8dB, respectively.

[0104] From the above simulations, it can be seen that the present invention proposes an adaptive block-based compressed sensing image reconstruction method based on texture information. Regardless of whether it is at a low sampling rate or a high sampling rate, the PSNR value and intuitive visual effect of the reconstructed image of the present invention are much higher than those of other existing algorithms, with an average improvement of 4.5-6.6dB; and it effectively suppresses the blocking effect, improves the subjective quality of image reconstruction and the peak signal-to-noise ratio while keeping the total sampling rate unchanged, and realizes a reasonable adaptive allocation of the sampling rate, which proves the effectiveness of the present invention.

[0105] The above detailed description of the preferred embodiments of the present invention does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the technical solution of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the technical solution of the present invention and shall remain within the scope of protection of the present invention.

Claims

1. An adaptive block-based compressed sensing image reconstruction method based on texture information, characterized in that: The following steps are involved: Step 1: Divide the original image into blocks; Step 2: extracting texture information of each image block and calculating its texture information weight, wherein the texture information includes grayscale entropy, standard deviation, and edge information; Step 3: clustering the image blocks according to the texture information weight differences of the image blocks calculated in step 2; Step 4: Calculate the adaptive sampling rate of each image block, construct a measurement matrix for each image block based on the adaptive sampling rate, and perform measurement. The specific contents of calculating the adaptive sampling rate of each image block include: Step 41): Set the total sampling rate SR∈[0.1,0.5] of the image; Step 42): Calculate the pre-sampling rate SR_Pre of each image block, the pre-sampling rate SR_Pre of image block j j Defined as: in are the texture information weights of each image block; W and H are the width and height of the original image respectively; B is the size of the image after block division; SR is the total sampling rate of the image; Step 43): According to the pre-sampling rate SR_Pre of each image block and the clustering result obtained in step 3, the pre-sampling of the same type of image blocks is first summed and then averaged to obtain the adaptive sampling rate SR of each type of image block. j ; Step 44): Adaptive sampling rate SR for each type of image block j Perform secondary distribution Set the sampling rate upper limit SR max , traverse each image block, if the sampling rate of an image block exceeds the upper bound of the sampling rate, calculate the part that exceeds the upper bound of the sampling rate: SR plus =SR plus +(SR j -SR max ) Among them SR max The upper limit of the sampling rate; after that, the part SR that exceeds the upper limit of the sampling rate max Evenly distribute to all image blocks. If there are still image blocks with sampling rates exceeding the upper limit of the sampling rate, repeat this step until all image blocks have SR j ≤SR max ; Step 5: Reconstruct the image using a smooth projection Landweber reconstruction algorithm based on a superposition structure block method to obtain large-size reconstructed image blocks for stitching. The specific process of the smooth projection Landweber reconstruction algorithm based on a superposition structure block method is as follows: Assume that the block sizes during measurement and reconstruction are B1 and B2 respectively, and to achieve the superposition structure, B1 and B2 satisfy the following relationship: B1<B2 and B2=2 l ×B1 l=1,2,3,...,L Where l is a positive integer. The superposition block structure first divides the two-dimensional image X into m non-overlapping blocks of size N during measurement. B =B1×B1 image block X bj (j=1,2,...,m), and transform it into a B ×1 one-dimensional vector x bj ; Then by constructing a B ×N B Measurement matrix Φ bj , where M B =SR j ×B is the total number of samples of each image block; the size of each block is M B ×1 measurement value vector y bj , which is expressed as follows: y bj =Φ bj ·x bj j=1,2,...,m At the reconstruction end, the measurement value vector y at the measurement end bj , the measurement matrix Φ bj and the one-dimensional vector x bj Splicing and reorganizing into large-size blocks; the reconstruction end sets l = 1, and the splicing process is as follows: where Φ ∧ are the four measurement matrices Φ bj The diagonal matrix formed by the concatenation of y and Φ ∧ The image is reconstructed as the input of the smooth projection Landweber reconstruction algorithm to obtain a large-size reconstructed image block.

2. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: The specific content of the block division of the original image in step 1 includes: dividing the image of size N=W×H into m image blocks of the same size of B×B and without overlapping, and the column vector form of the jth image block is x j , j=1,2,...,m,m=N / B 2 .

3. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: The content of calculating the texture information weight of each image block in the step 2 includes: dividing the image into blocks and calculating the grayscale entropy h of each image block j. j ,Std j , and edge information E j ; Then calculate the grayscale entropy, standard deviation and weight of edge information of each image block to the corresponding texture information of all image blocks, and get the grayscale entropy weight Standard deviation weight and edge weights 4. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: The specific process of clustering each image block in step 3 is as follows: the texture information weight of each image block calculated in step 2 is used as the input of the K-means++ clustering algorithm, and the K-means++ algorithm is used to cluster each image block according to the richness of texture information, and each image block is classified from less to more according to the richness of texture information:

5. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: The upper bound of the sampling rate is set to 0.

8.

6. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: The specific content of constructing the measurement matrix of various image blocks is: using a size of B 2 ×B 2 The orthogonal Gaussian random matrix is ​​used as the measurement matrix Φ; according to the adaptive sampling rate of the flat block, transition block and texture block, the number of measurements of each block is determined and their respective measurement matrices are generated. and Take measurements.

7. The adaptive block-based compressed sensing image reconstruction method based on texture information according to claim 1, characterized in that: It also includes judging whether the large-size reconstructed image block obtained in step five meets the peak signal-to-noise ratio requirement for the reconstructed image. If it meets the requirement, the image reconstruction process is terminated to obtain the reconstructed image, and all large-size reconstructed image blocks are spliced ​​to obtain the entire reconstructed image.

Images