Labyrinth information extraction method research based on compressed sensing

Abstract

The image processing technology is utilized to realize global feature extraction of labyrinth information, so that the task execution efficiency of the labyrinth robot can be effectively improved, and in order to reduce high calculation complexity and time complexity in labyrinth image processing, a compressed sensing theory is put forward to be applied to labyrinth information extraction. A labyrinth image adaptive sampling rate block compressed sensing (STA-ABCS-ASPL) algorithm based on statistical information is adopted, clustering analysis is carried out by using one-dimensional gray entropy and standard deviation of image blocks as prior information, and the image blocks are divided into simple blocks, smooth blocks and complex blocks. And adaptively pre-allocating a sampling rate to each image block according to the image statistical information, and adaptively allocating the sampling rate of each class of blocks in combination with a clustering result. When the image is reconstructed, the higher reconstruction quality is obtained by adopting an accelerated smooth projection Landweber algorithm (ASPL). A large number of maze image information extraction experiment results show that the PSNR performance of the STA-ABCS-ASPL algorithm is better improved compared with ABCS-SPL and ENT-ABCS-SPL algorithms, the global information of the maze can be extracted at the accuracy rate of 98.05%, and the method has practical application value.

Description

Technical Field

[0001] The present invention relates to a method for extracting maze information based on compressive sensing, belonging to the technical field of image processing. Background Art

[0002] With the changes of the times and the rapid development of intelligent technologies, mobile robots have increasingly become a popular technology in various fields in life and production. Using image processing technology to achieve global feature extraction of maze information can effectively improve the task execution efficiency of maze robots. However, the pixel level of maze images is in the millions, which occupies a large amount of CPU resources during the maze feature information extraction process, with relatively high computational complexity and time complexity. Compressive sensing reduces the dimension of image data during the image acquisition stage. Therefore, for large-scale image data processing, it can significantly reduce the amount of data processing and has strong adaptability and flexibility. Based on the above background, applying compressive sensing to maze image information extraction can effectively relieve the pressure on data storage space in image processing. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method for extracting maze information based on compressive sensing, adopting a statistical information-based adaptive sampling rate block compressive sensing (STA-ABCS-ASPL) algorithm for maze images, which can reduce the amount of data, lower the high computational complexity and time complexity in maze image processing, and improve the efficiency of maze image information extraction at the same time.

[0004] To achieve the above object, the present invention adopts the following specific scheme:

[0005] A method for extracting maze information based on compressive sensing, comprising the following steps:

[0006] Step S1: Perform perspective transformation preprocessing on the original maze images with different shooting angles, brightness, backgrounds, layouts, etc. obtained by a smart phone.

[0007] Step S2: Apply the statistical information-based adaptive sampling rate block compressive sensing algorithm (STA-ABCS-ASPL) for maze images to the preprocessed maze images, calculate the one-dimensional gray entropy H i , standard deviation σ i as prior information for clustering analysis, and divide the image blocks into simple blocks, flat blocks and complex blocks. Adaptive sampling rates are allocated to each image block according to the statistical information of the maze image, and the sampling rates of each type of block are adaptively allocated in combination with the clustering results:,, and a Gaussian random measurement matrix is used for observation.

[0008] Step S3: To improve the reconstruction quality and reconstruction speed of the maze image, the accelerated smoothed projection Landweber algorithm (ASPL) is used for image reconstruction.

[0009] Step S4: Perform edge detection on the reconstructed maze image using the Canny operator. This algorithm processes details more meticulously and coherently, and can depict the edges of the maze baffles more perfectly, providing a guarantee for subsequent accurate feature matching.

[0010] Step S5: Uniformly divide the maze edge contour map and identify the wall information of each maze cell for feature matching.

[0011] Furthermore, perform perspective transformation preprocessing on the original maze image. Due to the angle and distance relationship between the device and the maze during shooting, straight lines will appear as oblique lines in the picture. To correct and create a better visual effect, use perspective transformation to achieve distortion correction of the maze image and restore the true appearance of the maze image.

[0012] Furthermore, the specific content of step S2 is as follows:

[0013] Step S21: Let x i be the i-th image sub-block (i = 1, 2,..., J), and each image block x i is measured by the observation matrix Φ B to obtain yi = Φ B ·x j . Among them, Φ B is a measurement matrix of size M B ×B 2 , Therefore, the equivalent measurement matrix Φ of the entire image is a block diagonal matrix:

[0014] Step S22: Calculate the statistical information of the image block, the one-dimensional gray entropy H i , the standard deviation σ i and the edge information E i . Among them, P j is the proportion of pixels with gray value j in the image, B is the size of the image block x i , x i，j represents the pixel value of the image block x i at the position (i, j), represents the average gray value of the image block x i .

[0015] Step S23: Use the one-dimensional gray entropy and standard deviation of the maze image as the data set for clustering, and the maze image blocks can be divided into three categories: simple blocks, smooth blocks, and complex blocks.

[0016] Step S24: The one-dimensional gray entropy H of the image blocki , standard deviation σ i and edge information E i As input, the initial sampling rate R of each image block can be calculated i . where R is the total sampling rate, and R max is the upper limit value of the sampling rate, and satisfies the following relationship:

[0017] Step S25: According to the K-means clustering image block results and the initial sampling rate R of each image block i , the initial sampling rate R′ of the three types of image blocks is calculated by the following formula k (k = 1, 2, 3). In the formula, G is the number of each type of image block, and R kj is the initial sampling rate of each image block. Traverse the sampling rates of each type of block in turn. If R′ k ≥R max , sum all the parts greater than R max and evenly distribute them to the sampling rates less than R max , and set R′ k = R max to obtain the self-weighted adaptive sampling rate R′ k , and use the Gaussian random matrix for observation to obtain the measurement value y.

[0018] Furthermore, the specific steps of step S3 are as follows:

[0019] Input: Observation matrix Φ B , measurement value y, threshold parameter λ; Output: Reconstructed image x.

[0020] Step S31: Initialize x[0], set the maximum number of iterations k max , determine the value of γ, and calculate the convergence coefficient μ k and ω k .

[0021] Step S32: Use Wiener filtering to perform smoothing filtering on the image x[k] to obtain

[0022] Step S33: Perform improved projection Landweber iteration on each image block x i to obtain

[0023] Step S34: Perform projection sparse transformation, process it using the Threshold threshold, and then perform inverse transformation.

[0024] Step S35: Perform inverse transformation reconstruction:

[0025] Step S36: Let Repeat steps S33 to S35 for each image block.

[0026] Step S37: Judge the termination condition. When |D (k+1) -D (k) |< ε, stop the iteration.

[0027] Furthermore, for the canny edge detection processing of the reconstructed maze image, in order to obtain the wall information of all cells in the maze, it is necessary to detect the edges of the maze baffles, which can not only greatly reduce the data volume of maze image analysis and processing, but also facilitate the subsequent extraction of maze information. The Canny operator is used for edge detection of the maze map, and this algorithm processes details more carefully and coherently, and can depict the edges of the maze baffles more perfectly.

[0028] Furthermore, perform feature matching on the maze contour map after edge detection. The present invention extracts information based on a 16×16 maze map. In order to accurately obtain the information of maze cells, the obtained maze edge contour map is evenly divided into 256 blocks, so that each block of maze cells contains the information of the four surrounding walls of up, down, left, and right. By representing the state of each maze cell through the wall information, the information of the entire maze image can be obtained.

[0029] The present invention has the following advantages and positive effects:

[0030] First: The STA-ABCS-ASPL algorithm adopted by the present invention realizes adaptive sampling rate allocation for each type of image block. Since the sampling rates of various types of image blocks are fixed, only three observation matrices are required for image compressive sensing, thus reducing the occupation of storage space by the observation matrices.

[0031] Second: The accelerated smoothed projection Landweber algorithm is used for maze image reconstruction, which improves the quality of maze image reconstruction while reducing the reconstruction time consumption and improving the system operation efficiency.

[0032] Third: Applying compressive sensing to maze image processing effectively solves the problem of large data volume in maze image storage and transmission, and the accuracy of maze information extraction is relatively high, providing guarantee for subsequent path planning work. Description of the Drawings

[0033] Figure 1 is the system flow block diagram;

[0034] Figure 2 is the STA-ABCS-ASPL system framework diagram;

[0035] Figure 3 is the comparison of the PSNR performance of the maze image with the comparison algorithms (the comparison algorithms are ABCS-SPL and ENT--ABCS--SPL) at different sampling rates;

[0036] Figure 4 is the comparison of the SSIM performance of the maze image with the comparison algorithms (the comparison algorithms are ABCS-SPL and ENT--ABCS--SPL) at different sampling rates;

[0037] Figure 5 is the comparison of the reconstruction time performance of the maze image with the comparison algorithms (the comparison algorithms are ABCS-SPL and ENT-ABCS-SPL) at different sampling rates; Specific implementation manners

[0038] In order to make the technical solution of the present invention more clearly reflected, the technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation manners:

[0039] As Figure 1 shown is a system of a maze information extraction method based on compressive sensing. The system includes five links: maze image preprocessing, compressive sensing processing (STA-ABCS-APSPL algorithm), canny edge detection, and feature matching. Each link complements each other and extracts the global information of the maze with high accuracy

[0040] In the present invention, the maze information extraction method based on compressive sensing includes the following steps:

[0041] Step S1: Perform perspective transformation preprocessing on the original maze images with different shooting angles, brightness, backgrounds, layouts, etc. obtained by a smartphone;

[0042] Further, the system block diagram of the compressive sensing algorithm (STA-ABCS-ASPL) is as Figure 2 shown: Use the one-dimensional gray entropy and standard deviation of image blocks as prior information for clustering analysis, and divide the image blocks into simple blocks, flat blocks, and complex blocks. Adaptive pre-allocate the sampling rate for each image block according to the image statistical information, and adaptively allocate the sampling rate for each type of block in combination with the clustering result. When reconstructing the image, the accelerated smoothed projection Landweber algorithm (ASPL) is used to obtain a high reconstruction quality.

[0043] Further, the specific step S2 is:

[0044] Step S21: Let x i be the i-th image sub-block (i = 1, 2, …, J). Each image block x i is measured by the measurement matrix Φ B to obtain y i = Φ B ·x i . Where Φ B is a measurement matrix of size M B ×B 2 . Therefore, the equivalent measurement matrix Φ of the entire image is a block diagonal matrix:

[0045] Step S22: Calculate the statistical information of the image block, the one-dimensional gray entropy H i , the standard deviation σ i and the edge information E i . Where P j is the proportion of pixels with gray value j in the image, B is the size of the image block x i , x i，j represents the pixel value of the image block x i at the position (i, j), represents the average gray value of the image block x i .

[0046] Step S23: Using the one-dimensional gray entropy and standard deviation of the maze image as the data set for clustering, the maze image blocks can be divided into three categories: simple blocks, flat blocks, and complex blocks.

[0047] Step S24: Taking the one-dimensional gray entropy H i , the standard deviation σ i and the edge information E i of the image block as the input, the initial sampling rate R i of each image block can be calculated. Where R is the total sampling rate, R max is the upper limit value of the sampling rate, and the following relationship is satisfied:

[0048] Step S25: According to the K-means clustering image block result and the initial sampling rate R i of each image block, the initial sampling rate R′ k of the three types of image blocks is calculated by the following formula (k = 1, 2, 3). Where G is the number of various image blocks, and R kj is the initial sampling rate of each image block. Traverse the sampling rates of each type of block in sequence. If R′ k ≥R max , sum all the parts greater than R max and evenly distribute them to the sampling rates less than R max , and set R′ k =R max to obtain the self-weighted adaptive sampling rate R′ k , and use the Gaussian random matrix for observation to obtain the measurement value y.

[0049] Further, the specific steps of step S3 are as follows:

[0050] Input: Observation matrix Φ B , measurement value y, threshold parameter λ; Output: Reconstructed image x

[0051] Step S31: Initialize x [0] , set the maximum number of iterations k max , determine the value of Y, and calculate the convergence coefficients μ k and ω k

[0052] Step S32: Use Wiener filtering to perform smoothing filtering on the image x [k] to obtain

[0053] Step S33: Perform improved projection Landweber iteration on each image block x j to obtain

[0054] Step S34: Perform projection sparse transformation, and use Threshold threshold processing and then perform inverse transformation

[0055] Step S35: Perform inverse transformation reconstruction:

[0056] Step S36: Repeat steps S33 to S35 for each image block

[0057] Step S37: Judge the termination condition. When |D (k+1) -D (k) |<ε, stop the iteration

[0058] Further, Figure 3Comparison of the STA-ABCS-ASPL algorithm with the comparative algorithms ABCS-SPL and ENT-ABCS-SPL in terms of peak signal-to-noise ratio (PSNR) performance (SR = [0.1, 0.5]). Compared with the comparative algorithms, the PSNR value of the STA-ABCS-ASPL algorithm shows an increasing trend with the increase of the sampling rate, and the PSNR of this algorithm is better than that of the comparative algorithms at each sampling rate, with an average improvement of 3.168 dB and 2.086 dB, and the maze image reconstruction quality is better.

[0059] Furthermore, Figure 4 Comparison of the STA-ABCS-ASPL algorithm with the comparative algorithms ABCS-SPL and ENT-ABCS-SPL in terms of structural similarity (SSIM) performance (SR = [0.1, 0.5]). Compared with the comparative algorithms, the SSIM value of the STA-ABCS-ASPL algorithm also shows an increasing trend with the increase of the sampling rate, and the SSIM value of this algorithm is better than that of the comparative algorithms with an average improvement of 0.116 and 0.088.

[0060] Furthermore, Figure 5 Comparison of the STA-ABCS-ASPL algorithm with the comparative algorithms ABCS-SPL and ENT-ABCS-SPL in terms of reconstruction time consumption performance (SR = [0.1, 0.5]). The reconstruction time consumption of this algorithm is significantly shorter, which can improve the efficiency of the whole system.

[0061] Furthermore, for the canny edge detection processing of the reconstructed maze image, in order to obtain the wall information of all cells in the maze, it is necessary to detect the edges of the maze baffles, which can not only greatly reduce the data volume of maze image analysis and processing, but also facilitate the subsequent extraction of maze information. The Canny operator is used for edge detection of the maze map, and this algorithm processes details more carefully and coherently, and can depict the edges of the maze baffles more perfectly.

[0062] Furthermore, feature matching is performed on the maze contour map after edge detection. The present invention extracts information based on a 16×16 maze map. In order to accurately obtain the information of maze cells, the obtained maze edge contour map is evenly divided into 256 pieces, so that each piece of maze cell contains the information of the four surrounding walls of up, down, left and right. By representing the state of each maze cell through the wall information, the information of the entire maze image can be obtained.

Claims

1. A maze information extraction method based on compressed sensing, characterized in that include: Step S1: Preprocessing the original maze images with different shooting angles, brightness, backgrounds, layouts, etc. taken by the smartphone with perspective transformation; Step S2: Apply the maze image adaptive sampling rate block compressed sensing algorithm based on statistical information (STA-ABCS-ASPL) to the preprocessed maze image to calculate the one-dimensional grayscale entropy H of the maze image i , standard deviation σ i Cluster analysis is performed as prior information to divide image blocks into simple blocks, smooth blocks, and complex blocks. The sampling rate is adaptively allocated to each image block according to the maze image statistics, and the sampling rate of each type of block is adaptively allocated based on the clustering results: and a Gaussian random measurement matrix is ​​used for observation. Step S3: To improve the quality and speed of maze image reconstruction, the accelerated smooth projection Landweber algorithm (ASPL) is used to reconstruct the image. Step S4: The reconstructed maze image is subjected to edge detection using the Canny operator. This algorithm handles details more carefully and coherently, and can depict the edge of the maze baffle more perfectly, thus providing a guarantee for subsequent accurate feature matching. Step S5: Divide the maze edge contour map into blocks evenly, identify the wall information of each maze grid and perform feature matching.

2. The method for extracting maze information based on compressed sensing according to claim 1, characterized in that: The original maze image is preprocessed with perspective transformation. Due to the angle and distance relationship between the device and the maze during shooting, straight lines will appear as oblique lines in the picture. In order to correct and create a better visual effect, perspective transformation is used to achieve distortion correction of the maze image and restore the true appearance of the maze image.

3. The method for extracting maze information based on compressed sensing according to claim 1, characterized in that: The specific location of step S2 is: Step S21: Let x i is the i-th image sub-block (i=1,2,…,J), each image block x i From the observation matrix Φ B Measured y i =Φ B ·x i . Where Φ B The size is M B ×B 2 The measurement matrix, Therefore, the equivalent measurement matrix Φ of the entire image is a block-by-block diagonal matrix: Step S22: Calculate the statistical information of the image block, the one-dimensional grayscale entropy H i , standard deviation σ i And edge information E i . Where P j is the proportion of pixels with grayscale value j in the image, and B is the image block x i The size of x i,j Represents image block x i The pixel value at position (i, j), Represents image block x i The average gray value. Step S23: using the one-dimensional grayscale entropy and standard deviation of the maze image as a clustering data set, the maze image blocks can be divided into three categories: simple blocks, smooth blocks and complex blocks. Step S24: The one-dimensional grayscale entropy H of the image block i , standard deviation σ i And edge information E i As input, the initial sampling rate R of each image block can be calculated i Where R is the total sampling rate, R max is the upper limit of the sampling rate and satisfies the following relationship: Step S25: Based on the K-means clustering image segmentation result and the initial sampling rate R of each image block i , the initial sampling rate R′ of the three types of image blocks is calculated by the following formula k (k=1, 2, 3). Where G is the number of each type of image blocks, R kj is the initial sampling rate of each image block. Traverse the sampling rate of each type of block in turn. If R′ k ≥R max , all the values ​​greater than R max The partial sum is evenly distributed to the max The sampling rate of R′ k =R max , and obtain the adaptive sampling rate R′ of the self-weight k , and use the Gaussian random matrix to observe and obtain the measured value y.

4. The method for extracting maze information based on compressed sensing according to claim 1, characterized in that: The step S3 is specifically as follows: Input: observation matrix Φ B , measurement value y, threshold parameter λ. Output: reconstructed image x. Step S31: Initialize x [0] , set the maximum number of iterations k max , determine the value of γ and calculate the convergence coefficient μ k and ω k . Step S32: Use Wiener filtering to process the image x [k] Perform smoothing filtering to obtain Step S33: For each image block x, i Improved projection Landweber iteration is performed to obtain Step S34: Perform sparse transformation on the projection, use Threshold threshold processing and then perform inverse transformation. Step S35: Perform inverse transformation and reconstruction: Step S36: Repeat steps S33 to S35 for each image block. Step S37: Determine the termination condition. (k+1) -D (k) When |<ε, stop iterating.

5. The method for extracting maze information based on compressed sensing according to claim 1, characterized in that: The reconstructed maze image is subjected to canny edge detection. In order to obtain the wall information of all cells in the maze, the edge of the maze baffle needs to be detected, which can not only greatly reduce the amount of data for maze image analysis and processing, but also facilitate the subsequent extraction of maze information. The edge detection of the maze map is performed using the Canny operator. The algorithm handles details more carefully and coherently, and can perfectly depict the edge of the maze baffle.

6. The method for extracting maze information based on compressed sensing according to claim 1, characterized in that: The present invention performs feature matching on the maze outline map after edge detection. The present invention extracts information based on a 16×16 maze map. In order to accurately obtain the information of the maze cells, the obtained maze edge outline map is evenly divided into 256 blocks, so that each maze cell contains the information of the four walls on the top, bottom, left and right sides. The state of each maze cell is represented by the wall information, and the information of the entire maze image can be obtained.

Images