A rock mass fracture identification method

Abstract

The application discloses a rock mass crack identification method, which belongs to the technical field of rock mass engineering and comprises the following steps: firstly, traversing an image to obtain extreme points of a gray function of a normal direction of a pixel point, i.e. coarsely extracting a crack center line; calculating a score of each item based on Hessian eigenvalue, symmetry and gray value respectively and calculating a confidence degree by weighted average; then, further screening candidate points by threshold limitation and non-maximum suppression; next, connecting the candidate points and discontinuous cracks to form a complete crack center line; and finally, removing pseudo-cracks and noises by calculating average gray and length of each connecting line to obtain a complete and accurate crack center line. The application is not only suitable for simple concrete cracks, but also can achieve satisfactory effects in complex rock mass crack detection, and solves the problem of identifying cracks by a computer in a rock mass with complex composition and many interferences.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of rock mass engineering, and particularly relates to a rock mass fracture identification method. BACKGROUND

[0002] With the increasing development of national infrastructure construction and underground space, rock mass engineering such as open pit slope, underground powerhouse, subway and tunnel through mountains is gradually becoming a widely existing construction project. In the process of project construction, the quality of rock mass can have a significant impact on design, construction scheme, budget and construction period. The condition of rock mass fracture joint is an important indicator for evaluating the quality of rock mass. Joint and fracture can cause damage, deformation and sliding of rock mass, thereby negatively affecting the stability of rock structure and the safety of rock engineering. Therefore, accurate identification and extraction of fractures in rock mass are of great significance to the design, construction and evaluation of rock mass engineering.

[0003] The traditional rock mass fracture identification method is mainly direct visual method. First, the surface of the geological body is visually observed to find the fracture, then the rock mass is recorded by sketching, and finally the quality of surrounding rock is evaluated by experts on site according to experience. The traditional method is simple and practical, but the identification and measurement of fractures by artificial experience judgment have subjectivity and large error; and the geological sketching consumes a lot of manpower and time cost. Therefore, more and more engineering and scientific research personnel use machine vision to detect the surrounding rock fracture joint. The fracture identification method based on machine vision mainly includes image processing and neural network. The first kind is to perform binarization, filtering and other operations on the fracture through image processing, and finally separate the background and the fracture. The other kind is to use machine learning to artificially mark the fracture image to obtain a fracture training set. Then a neural network model is built, and the model can better identify the fracture in the rock mass image through training. The traditional method of identifying cracks using image processing can only achieve good results in the scene of concrete cracks. However, these fractures have the characteristics of simple structure, obvious contrast with background, small interference and less noise. When facing complex rock mass fractures, satisfactory results cannot be achieved. The use of machine learning neural network requires the construction of a complex neural network model. Moreover, the data set required for training generally needs artificial fracture labeling, and the data set production also consumes a lot of manpower and time cost. Therefore, how to accurately, efficiently and low-cost detect cracks in complex rock mass with many interferences is still a challenging problem. SUMMARY

[0004] In order to solve the above problems, the present application provides a rock mass fracture identification method, which comprises the following steps:

[0005] S1, image processing operation is performed on the crack image, and a crack center point is roughly extracted by traversing the image after the image processing;

[0006] S2, the crack center point after the rough extraction is evaluated for a characteristic value, and a confidence level is obtained by weighting and synthesizing;

[0007] S3, the confidence level of the candidate point is threshold limited and non-maximum value suppressed, and the candidate point is further filtered to facilitate the connection of the candidate point;

[0008] S4, the candidate points are connected to form a crack center connection line, and the discontinuous crack connection lines are further connected;

[0009] S5, the pseudo-cracks are removed to obtain a crack center connection line with high precision.

[0010] Further, the image processing operation includes: converting the image into a gray image, performing median filtering and Gaussian filtering; the crack center point rough extraction step of the image includes:

[0011] First, the Hessian matrix of each pixel point is obtained by traversing the image, the image gray of the crack center line normal direction is approximately subject to Gaussian distribution, assuming that the gray value of a point (x0, y0) on the image is I(x0, y0), and I(x, y) is the gray distribution function on the two-dimensional image, which is a binary function; the n-order function is expanded by Taylor at x k , and the part above the second order is discarded to write in the form of a matrix:

[0012]

[0013] The Hessian matrix is expressed as:

[0014]

[0015] In the formula: I xx , I xy , I yy are obtained by convolution of the Gaussian convolution kernel and the image gray I(x, y);

[0016] Further, for a point (x0, y0) on the image, the gray distribution function I(x0, y0) in the normal direction is expanded by Taylor along the direction of the point, and the unit direction vector in the normal direction is e=(n x , n y ), then the gray value of any point in the normal direction is expressed as:

[0017]

[0018] Wherein: I(x0, y0) and direction vector e are calculated by image information; the above formula is regarded as a unary function about t, and the value of t is solved by taking the first derivative of t as zero:

[0019]

[0020] Further, if It is explained that the extreme point is in the current pixel point, and the current pixel point (x0, y0) is regarded as the extreme point of the gray distribution, and the point meeting the condition is regarded as the crack center point after rough extraction.

[0021] Further, the characteristic value evaluation of the crack center point after rough extraction includes three aspects of Hessian matrix-based characteristic value evaluation, crack normal direction symmetry evaluation and gray value evaluation;

[0022] The Hessian matrix-based characteristic value evaluation method is to calculate two characteristic values λ1, λ2 of the Hessian matrix of each pixel point, and the calculation formula is:

[0023]

[0024] Wherein: the calculation methods of K and Q are:

[0025]

[0026] The Hessian matrix characteristic values of the actual crack center point should satisfy:

[0027] λ1≈0, |λ1|<<λ2

[0028] For the crack center point that has been roughly selected, the size of the Hessian matrix characteristic value is calculated one by one, and the point is scored according to the degree of meeting the above condition, the higher the score of the point meeting the condition, and the full score is 100;

[0029] The crack normal direction symmetry evaluation method is:

[0030] Let the gray distribution functions of the image at the left and right symmetric positions of the candidate point (x0, y0) be f(x i , y i ), g(x j , y j ), wherein (x i , y i ), (x j , y j ) are symmetric points about (x0, y0), and the gray distribution functions on both sides are written as:

[0031] f(x, y) = I(x0+tn x, y0 + tn y ),

[0032] g(x, y) = I(x0 - tn x , y0 - tn y )

[0033] The pixel points on the image are discontinuous. In order to measure the degree of symmetry, the following is defined:

[0034] S i 2 = [f(x i , y i )-g(x j , y j )] 2

[0035] In the formula: it is the square of the gray value difference of the point symmetrical about (x0, y0), t is in a certain range, t ∈ [0, T], and the symmetry degree is expressed as:

[0036]

[0037] In the formula: T represents the number of pixels calculated, each time S is calculated, one pixel point is advanced, and the number of calculated pixel points is set by setting the value of T; according to the size of S at the point (x0, y0), the normal direction gray value symmetry of the candidate point is evaluated, the smaller S is, the higher the symmetry degree is, and the higher the score is, and the full score is 100 points;

[0038] The gray value-based feature value evaluation method is:

[0039] The gray value in the image is mapped into a gray score value in the range of 0-100; wherein the score value of the gray value less than 25 is 100, the gray value of the pixel point 25-125 is linearly mapped into the score value of 100-0, and the score of the pixel point higher than 125 is 0;

[0040] After calculating each item score of each point according to the above method, the score is adjusted, and the specific method is:

[0041] The first ranked score is the highest, which is 100, and the score is sequentially decreased according to the ranking, and the score of the last ranked is 0. Each item is adjusted and then weighted and averaged, and finally the final score is adjusted again according to the above method. The weighted and averaged score is divided by 100 to obtain the confidence of the candidate point.

[0042] Further, in the step S3, threshold limiting is first performed, and points with a confidence (0-1) lower than a certain threshold are removed, and non-maximum suppression is performed on the remaining points; the non-maximum suppression includes two aspects, which are feature vector direction non-maximum suppression and confidence non-maximum suppression;

[0043] First, the step of feature vector non-maximum suppression is: first, the directions around a point are divided into 6 regions, and points within a certain range of the point are divided into 6 categories according to the regions where the directions of the feature vectors e1 of the points point to; the points in each region are a category, and within a certain range of the point, only the points whose feature vectors point to the most regions are retained by using the non-maximum suppression NMS method;

[0044] Second, the step of confidence non-maximum suppression is: by selecting a candidate point in the neighborhood of the candidate point, a candidate point with a higher score is obtained, and a candidate point with a lower score is discarded; thus, a candidate point that is most likely to be a crack center point is obtained, thereby sparsifying the candidate points, reducing the candidate points that are useless for connecting cracks, and improving the connection accuracy of the candidate points.

[0045] Further, in the step S4, the step of connecting the candidate points is:

[0046] The candidate points are sorted in descending order of confidence, and the search step and range are set from the first point, and the candidate points are searched along the tangent direction e1 and the opposite direction thereof; if the next candidate point is not the first and second two points of the straight line, the angle formed by the three connected lines is calculated, and if the angle is greater than a certain threshold value a and the difference between the gray value of the next candidate point and the gray value of the previous point is less than a certain threshold value b, the next candidate point and the previous candidate point are regarded as the center points of the same crack, and are connected; the above steps are repeated to connect the candidate points on each crack point by point, until the search range is within the threshold value or there is no candidate point to connect within a certain range of the last candidate point, and then the crack is regarded as reaching the end.

[0047] The connection step of the broken cracks is: starting from the first crack, the distances between the head and tail points of other cracks and the head and tail points of the crack are calculated in sequence to obtain the two points with the shortest distance; first, the angle formed by the two-point extension line and the connected line of the adjacent point on the straight line is calculated; second, the distance between the two points is calculated; Length_threshlod and Angle_threshlod represent the threshold value of the distance between the end points and the threshold value of the angle formed by the two adjacent crack end points and the crack point adjacent to one of the end points; the crack end points less than the threshold values Length_threshlod and Angle_threshlod are connected; the merged cracks are regarded as the last, and then the calculation and connection are performed from the next crack, until all the cracks do not meet the connection conditions, and the connection of the broken cracks is completed.

[0048] Further, the method for removing the pseudo cracks is:

[0049] In the ith crack, the distance between the two nodes is L ij , which is calculated by the following formula:

[0050]

[0051] In the formula: L i Length of the i-th crack; the length of each crack is obtained by the above formula; the cracks with smaller length are directly removed compared with the longest crack in the structure; in addition, for the complex crack situation of the rock mass, there are many unexpected noise and pseudo-cracks; although the previous steps cannot remove them, since they are not real cracks, the gray value of the pixel points they pass through is much larger than the gray value of the real crack position; these pseudo-cracks are removed by calculating the average gray value, and the calculation method of the average gray value is as follows:

[0052] GV i = (∑gv x，y ) / L i

[0053] In the formula: GV i represents the average gray value of the i-th crack, gv x，y represents the gray value of all pixel points passed by the line; the pseudo-cracks higher than the threshold value are filtered out by setting the gray value threshold GV_threshold.

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

[0055] (1) Compared with the traditional crack extraction method, the present application is safer and more cost-saving, does not need a large number of manually labeled crack pictures as a data set, does not need manual sketching to obtain the crack joint information of the rock mass, and can save a lot of manpower and resources.

[0056] (2) Wide applicability and high accuracy, whether it is a simple concrete crack or a rock mass under complex conditions of a real rock mass engineering, it has high accuracy and superiority.

[0057] (3) Good robustness and fast speed, which can efficiently remove the interference such as weeds and shadows around the crack rock mass, and compared with the crack recognition algorithm of the neural network, the time for making the data set and training is saved. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is the overall flow structure schematic diagram of the present application;

[0059] Figure 2 is the feature vector non-maximum suppression partition schematic diagram of the present application;

[0060] Figure 3 is the confidence non-maximum suppression method schematic diagram of the present application;

[0061] Figure 4 is the candidate point connection method schematic diagram of the present application;

[0062] Figure 5 An example effect diagram of the present application. DETAILED DESCRIPTION

[0063] In order to accurately, efficiently and at low cost identify cracks in complex rock mass images, the present application provides a rock mass crack identification method. The method extracts candidate points that meet the crack center point features based on confidence and non-maximum suppression, and further screens candidate points to obtain more likely candidate points, so as to connect the candidate points. Through further candidate point connection and pseudo-crack removal, a crack identification line with high precision is obtained.

[0064] The technical solution is as follows:

[0065] A rock mass crack identification method, the steps are as follows:

[0066] S1: performing image processing operation on the crack image, and traversing the crack center points after rough extraction through image processing.

[0067] S2: evaluating the eigenvalues of the obtained rough extraction crack center points, and weighting and synthesizing to obtain the confidence.

[0068] S3: threshold limiting and non-maximum suppression are performed on the confidence of the candidate points, and the candidate points are further filtered to facilitate the connection of the candidate points.

[0069] S4: connecting the candidate points to form a crack center connection line, and further connecting the discontinuous crack connection lines.

[0070] S5: removing pseudo-cracks to obtain a crack identification line (center line) with high precision.

[0071] In the S1, the image processing operation includes: converting the image into a grayscale image, performing median filtering and Gaussian filtering. The crack center point rough extraction step of the image includes:

[0072] First, traverse the image to obtain the Hessian matrix of each pixel point. Since the image gray value in the normal direction of the crack center line approximately obeys Gaussian distribution, it is assumed that the gray value of a point (x0, y0) on the image is I(x0, y0), and I(x, y) is the gray distribution function on the two-dimensional image. The distribution function is a binary function. The n-order function is expanded by Taylor at x k

[0073]

[0074] The Hessian matrix can be expressed as:

[0075]

[0076] ​In the formula: I xx , I xy , I yy are obtained by convolution of the Gaussian convolution kernel and the image gray scale I(x, y).

[0077] Further, for a point (x0, y0) on the image, the gray scale distribution function I(x0, y0) in the normal direction of the point is Taylor expanded along the direction of the point, and the unit directional vector in the normal direction is e=(n x , n y ), then the gray scale value of any point in the normal direction can be expressed as:

[0078]

[0079] In the formula: I(x0, y0) and the directional vector e can be calculated from the image information. Thus, the above formula can be regarded as a unary function about t, and the value of t can be solved by taking the first derivative of t and setting the first derivative to zero:

[0080]

[0081] Further, if It is explained that the extreme point is in the current pixel point, and the current pixel point (x0, y0) can be regarded as the extreme point of the gray scale distribution, and the points meeting this condition can be regarded as the center points of the cracks after rough extraction.

[0082] Preferably, the evaluation of the characteristic value of the center point of the crack after rough extraction in S2 includes three aspects of evaluation based on the eigenvalue of the Hessian matrix, evaluation of the symmetry of the normal direction of the crack, and evaluation of the gray scale value.

[0083] Preferably, the evaluation method based on the eigenvalue of the Hessian matrix is to calculate the two eigenvalues λ1 and λ2 of the Hessian matrix of each pixel point, and the calculation formula is:

[0084]

[0085] In the formula: the calculation methods of K and Q are:

[0086]

[0087] The eigenvalues of the Hessian matrix of the actual center point of the crack should satisfy:

[0088] λ1≈0, |λ1|<<λ2

[0089] For the center points of the cracks that have been roughly selected, the sizes of the eigenvalues of the Hessian matrix of each center point are calculated, and the center point is scored according to the degree of meeting the above condition, and the higher the score of the point, the more the point meets the condition, and the full score is 100.

[0090] Preferably, the method for evaluating the symmetry of the crack normal direction is as follows:

[0091] Let the gray-level distribution functions of the image at the positions symmetrical to the left and right of the candidate point (x0, y0) be f(x0, y0) respectively. i y i ), g(x j y j ), where (x i y i ), (x j y j Let (x0, y0) be a point symmetric about (x0, y0). Based on the t introduced above, the gray-level distribution function on both sides can be written as:

[0092] f(x, y) = I(x0 + tn) x y0+tn y ),

[0093] g(x, y) = I(x0 - tn) x y0-tn y )

[0094] Since the pixels in an image are not continuous, to measure the degree of symmetry, we define:

[0095] S i 2 =[f(x i y i )-g(x j y j )] 2

[0096] In the formula: represents the square of the difference in gray values ​​between points symmetric about (x0, y0). Within a certain range t∈[0, T], its degree of symmetry can be expressed as:

[0097]

[0098] In the formula: T represents the number of pixels to be calculated. Each calculation of S advances by one pixel. The number of pixels to be calculated can be set by setting the value of T. Based on the value of S at the point (x0, y0), the symmetry of the gray value in the normal direction of the candidate point can be evaluated. The smaller S is, the higher the symmetry and the higher the score. The maximum score is 100 points.

[0099] Preferably, the feature value evaluation method based on grayscale values ​​is as follows:

[0100] The grayscale values ​​in the image are mapped to grayscale scores ranging from 0 to 100. Grayscale values ​​less than 25 are scored as 100, and grayscale values ​​of pixels between 25 and 125 are linearly mapped to scores between 100 and 0. Pixels with grayscale values ​​greater than 125 are scored as 0.

[0101] Preferably, after calculating the scores for each point using the above method, adjustments are needed. Since the overall score distribution of the crack points in each image differs and is unlikely to be uniformly distributed between 0 and 100 points, a ranking-based scoring system is required to avoid the threshold needing to be reset due to image changes. The specific method is as follows:

[0102] The highest score is 100, decreasing sequentially with the lowest score being 0. Each item is adjusted before a weighted average is calculated. The final score is then adjusted again using the same method. This method applies not only to rock mass cracks but also to concrete cracks. For concrete cracks, the weight of the symmetry score should be appropriately increased, while for rock mass cracks, the weight of the symmetry score should be appropriately decreased. Dividing the weighted average score by 100 yields the confidence level of the candidate point.

[0103] Preferably, in step S3, a threshold constraint is first applied, removing points with a confidence level (0-1) below a certain threshold, and then non-maximum suppression is applied to the remaining points. Non-maximum suppression mainly includes two aspects: non-maximum suppression in the feature vector direction and non-maximum suppression based on confidence level.

[0104] First, the steps of eigenvector non-maximum suppression are as follows: First, the direction surrounding a point is divided into 6 regions. Points within a certain range of each point are classified into 6 classes based on the region pointed to by their eigenvector e1. Points in each region belong to one class. Within a given range, non-maximum suppression (NMS) is used to retain only points whose eigenvectors point to at most one region. Since the eigenvector e1 of a fracture point points to the fracture propagation direction, the direction pointed to by the tangential vector has absolute dominance in the local region of candidate points. This method can remove pseudo-fracture points with high confidence but whose tangential vectors point in other directions, thus facilitating the connection of fracture points.

[0105] Second, the steps of confidence non-maximum suppression are as follows: select candidate points with higher scores from the neighborhood of a candidate point and discard candidate points with lower scores. This obtains the candidate points that are most likely to be the center of the fracture, thereby sparsifying the candidate points, reducing the number of candidate points that are useless for fracture connection, and improving the accuracy of candidate point connection.

[0106] Preferably, in step S4, the step of connecting candidate points is as follows:

[0107] The candidate points are sorted by confidence from large to small. From the first point, the search step and range are set, and the candidate points are searched along the tangent direction e1 and the opposite direction. If the next candidate point is not the first and second points of the straight line, the angle formed by the three points is calculated. If the angle is greater than a certain threshold a and the difference between the gray value of the next candidate point and the gray value of the previous point is less than a certain threshold b, the next candidate point and the previous candidate point are regarded as the center points of the same crack, and are connected. The above steps are repeated to connect the candidate points on each crack point by point until the search range is reached. If the candidate point connection angle is less than the threshold or there is no candidate point within a certain range of the last candidate point, it is regarded as the end of the crack.

[0108] The connection step of the broken crack is as follows: starting from the first crack, the distances between the head and tail points of other cracks and the head and tail points of the crack are calculated in turn to obtain the two points with the shortest distance. First, the angle formed by the connecting line of the two points and the adjacent point on the straight line is calculated. Second, the distance between the two points is calculated. Length_threshlod and Angle_threshlod represent the threshold of the distance between the end points and the angle threshold formed by the two adjacent crack end points and the crack point adjacent to one of the end points. The crack end points less than the threshold Length_threshlod and Angle_threshlod are connected. The merged crack is regarded as the last, and then the calculation and connection are performed from the next crack until all cracks do not meet the connection condition, and the connection of the broken crack is completed.

[0109] Preferably, the method for removing the pseudo crack in S5 is:

[0110] In the i-th crack, the distance between the two nodes is L ij , which is calculated by the following formula:

[0111]

[0112] In the formula, L i represents the length of the i-th crack. The length of each crack can be obtained by the above formula. Compared with the longest crack in the structure, the cracks with smaller lengths can be directly removed. In addition, for the complex crack situation of the rock mass, there are many unexpected noise points and pseudo cracks. Although the previous steps cannot remove them, since they are not real cracks, the gray value of the pixel points is much larger than the gray value of the real crack position. These pseudo cracks can be removed by calculating the average gray value, and the calculation method of the average gray value is as follows:

[0113] GV i = (∑gv x，y ) / L i

[0114] In the formula, GVi Gv represents the average gray value of the i-th crack, gv x，y Gv represents the gray value of all pixel points through which the line passes. By setting a gray value threshold GV_threshold, false cracks higher than the threshold can be filtered out.

[0115] A rock mass crack identification method, as shown in Figure 1 The flow of the present application is to roughly extract crack center points, then evaluate the features of the roughly extracted points, including eigenvalue evaluation of the Hessian matrix, symmetry evaluation of the crack normal direction, and gray value evaluation. The evaluated values are integrated and threshold limited. Further, filtered candidate points are obtained. In order to facilitate the connection of the candidate points, non-maximum suppression is performed, including non-maximum suppression of the feature vector direction and non-maximum suppression of the confidence. Further, the candidate points are connected and non-continuous connection lines are filtered out to obtain crack identification lines with high accuracy

[0116] In this example, first, a rock mass image in which cracks are to be identified is acquired by a camera. First, image processing is performed on the image, including median filtering and Gaussian filtering. Median filtering can reduce the impact of uneven surfaces and shadows, and will not affect the image geometric information at the crack. The image after Gaussian filtering can be convolved by the following formula:

[0117]

[0118] The Hessian matrix of each pixel point is obtained by the above formula, and the formula is:

[0119]

[0120]

[0121] The two eigenvalues λ1, λ2 of the Hessian matrix can be calculated, and the points that meet the conditions are selected as the roughly extracted points, and the conditions are as follows:

[0122] λ1≈0, |λ1|<<λ2

[0123] Where < < is generally set to the condition that λ2 is more than ten times λ1. For λ1≈0, |λ1|<0.5 is generally set. For the candidate points that have been roughly extracted, the eigenvalue score of the Hessian matrix is evaluated, and the score is 0-100.

[0124] Preferably, the crack normal direction symmetry evaluation is performed. For the points after rough extraction, there are two Hessian matrix eigenvectors, and the direction with the fast falling gray gradient is the normal direction. For the crack structure, the gray on both sides of the crack is generally symmetrically distributed about the crack center line. T (generally set to 10) points in the normal direction of the center point of the rough extraction are selected, and the asymmetry is calculated, which is denoted by S, and the calculation formula is as follows:

[0125]

[0126] In the formula, T represents the number of pixels calculated. The greater the value of S, the lower the symmetry, and vice versa. The symmetry is scored, and the score is also between 0 and 100. Through this method, the interference of the edges of the detected object as the crack points can be avoided.

[0127] Preferably, the gray value-based eigenvalue evaluation is performed, and the gray value in the image is mapped into a gray score value in the range of 0-100. The score of the gray value less than 25 is 100, the gray value of the pixel point 25-125 is linearly mapped into the score of 100-0, and the score of the pixel point higher than 125 is 0.

[0128] Preferably, the values after the evaluation of the three methods are weighted and averaged. If the concrete crack is detected, the proportion of the symmetry score needs to be adjusted to be higher; if the rock mass crack is detected, the proportion of the symmetry score needs to be adjusted to be lower. In this embodiment, the proportions of the Hessian matrix eigenvalue evaluation, the symmetry evaluation, and the gray value evaluation are 40%, 30%, and 30%, respectively.

[0129] Preferably, the scores after the comprehensive evaluation are sorted from large to small, and the scores are given according to the ranking, so that the uneven distribution of the scores of each crack image can be avoided. Specifically, the score of the first ranking is the highest, which is 100, and the score of the last ranking is 0. Each item is adjusted and then weighted and averaged. Finally, the final score is adjusted again according to the above method. This method is not only suitable for rock cracks, but also suitable for concrete cracks. For the concrete cracks, the proportion of the symmetry score should be appropriately adjusted to be higher, and for the rock cracks, the proportion of the symmetry score should be appropriately adjusted to be lower. The final score is divided by 100 to obtain a value between 0 and 1, which is regarded as the confidence of each point as a crack center point.

[0130] Preferably, the obtained confidence is compared with the set confidence threshold (generally set to 0.8), and the rough extraction points lower than the set confidence threshold are filtered out to obtain candidate points.

[0131] Preferably, the non-maximum suppression of the feature vector is performed on the obtained candidate points. First, as shown in FIG. 6, the feature vector of each candidate point is calculated, and the feature vector of each candidate point is compared with the feature vectors of the adjacent candidate points. Figure 2The plane is divided into 6 regions, and due to the interference of shadows and irregular rock surfaces, the tangential direction of each candidate point is not necessarily strictly directed to the direction of the crack propagation. In order to facilitate the connection of candidate points, the point set in the vicinity of each point is divided into 6 categories according to the direction indicated by the tangential feature vector, and the points in the non-maximum number of regions are removed to facilitate the connection of subsequent candidate points. Further, the processed points are subjected to confidence non-maximum suppression, as shown in the figure, and the points with the maximum confidence in the local range are retained to ensure that multiple points available for connection in the local area affect the connection of the candidate points. Figure 2 The plane is divided into 6 regions, and due to the interference of shadows and irregular rock surfaces, the tangential direction of each candidate point is not necessarily strictly directed to the direction of the crack propagation. In order to facilitate the connection of candidate points, the point set in the vicinity of each point is divided into 6 categories according to the direction indicated by the tangential feature vector, and the points in the non-maximum number of regions are removed to facilitate the connection of subsequent candidate points. Further, the processed points are subjected to confidence non-maximum suppression, as shown in the figure, and the points with the maximum confidence in the local range are retained to ensure that multiple points available for connection in the local area affect the connection of the candidate points. Figure 3 The plane is divided into 6 regions, and due to the interference of shadows and irregular rock surfaces, the tangential direction of each candidate point is not necessarily strictly directed to the direction of the crack propagation. In order to facilitate the connection of candidate points, the point set in the vicinity of each point is divided into 6 categories according to the direction indicated by the tangential feature vector, and the points in the non-maximum number of regions are removed to facilitate the connection of subsequent candidate points. Further, the processed points are subjected to confidence non-maximum suppression, as shown in the figure, and the points with the maximum confidence in the local range are retained to ensure that multiple points available for connection in the local area affect the connection of the candidate points.

[0132] Preferably, the search step and range are set to search for candidate points along the tangential direction e1 and the opposite direction. If the next candidate point is not the first and second points of the straight line, the angle formed by the three-point line is calculated, and if the angle is greater than a certain threshold value a and the gray value difference between the next candidate point and the previous point is less than a certain threshold value b, the next candidate point and the previous candidate point are regarded as the center points of the same crack, and are connected, and the above steps are repeated.

[0133] Further, each candidate point on each crack point is connected in turn until the search range has candidate points connected by an angle less than a threshold value or the last candidate point has no candidate points connected within a certain range, and then the crack is considered to have reached the end. Repeat the above steps until all cracks are connected. Then connect the disconnected crack connection lines, as shown in the example effect diagram Figure 5 As shown in the figure, for the two end points with a distance less than a certain threshold value Length_threshlod and an included angle less than a certain threshold value Angle_threshlod, the two adjacent end points are connected, and finally a complete crack is formed. Figure 4

[0134] Preferably, the obtained crack connection lines are screened to filter out false cracks, and the specific removal method is to calculate the length of the crack connection line, and for the crack connection line with a length less than the maximum crack length, it is removed. The crack length calculation formula is as follows:

[0135]

[0136] In the formula, L i represents the length of the i-th crack, and the distance between the two nodes is L ij . Further, a better method to identify false cracks is to calculate their average gray value, and the calculation formula is as follows:

[0137] GV i = (∑gv x，y ) / L i

[0138] In the formula, GV​i Gv represents the average gray value of the i-th fissure, gv x，y Gv represents the gray value of all pixels through which the line passes. By setting a gray value threshold GV_threshold, false fissures higher than the threshold can be filtered out.

[0139] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacements or changes within the technical range disclosed in the present application according to the technical solution and concept of the present application, which should be covered in the protection scope of the present application.

Claims

1. A method of identifying fractures in a rock mass, characterized by, It comprises the following steps: S1, image processing operation is carried out on the crack image, and the crack center points are roughly extracted by traversing the image after the image processing; S2, the feature value of the obtained rough extraction crack center point is evaluated, and the confidence is obtained by weighted synthesis; S3, the confidence of the candidate point is threshold limited and non-maximum suppression, and the candidate point is further filtered to facilitate the connection of the candidate point; S4, the crack center connecting line is connected by connecting the candidate points, and the discontinuous crack connecting line is further connected; S5, the pseudo crack is removed to obtain a crack center connecting line with high precision; In the step S1, the processing operation of the image comprises: converting the image into a gray image, performing median filtering and Gaussian filtering; the crack center point rough extraction step of the image comprises: First, traverse the image to obtain the Hessian matrix of each pixel point. The image gray value in the normal direction of the center line of the crack approximately obeys the Gaussian distribution. Assuming that the gray value of a point on the image is is the gray distribution function on the two-dimensional image, which is a binary function; the n-ary function is Taylor expanded at , and the part higher than the second order is abandoned, which is written in the matrix form as:​ The Hessian matrix is expressed as: In the formula: , , All are obtained by convolving the Gaussian convolution kernel with the image gray scale ​ Further, for a certain point on the image the gray scale distribution function along the normal direction of which The Taylor expansion along the direction of the point is made, assuming that the unit directional vector of the normal direction is Then the gray scale value of any certain point on the normal direction is expressed as: wherein: and the direction vector are calculated from the image information; the above equation is considered as a unary function with respect to t, and the value of t is solved by taking the derivative of t and setting the first derivative to zero: Further, if the extreme point is within the current pixel point, the current pixel point is regarded as an extreme point of the gray scale distribution, and the point satisfying this condition is regarded as a center point of the crack after rough extraction.

2. The method of claim 1, wherein, In the step S2, the feature value evaluation of the rough extraction crack center point comprises: three aspects of Hessian matrix feature value evaluation, crack normal direction symmetry evaluation and gray value evaluation; The eigenvalue evaluation method based on the Hessian matrix is to calculate two eigenvalues of the Hessian matrix of each pixel point , , and the calculation formula is: The calculation method of K and Q is: The Hessian matrix feature value of the actual crack center point should satisfy: For the crack center points that have been roughly selected, the size of the Hessian matrix feature value is calculated one by one, and the points are scored according to the degree of compliance with the above conditions, the points with higher compliance with the conditions are scored higher, and the full score is 100 points; The crack normal direction symmetry evaluation method is: Let its image at the candidate point The gray level distribution functions at the symmetric positions are respectively where , is about the symmetric point, the gray level distribution functions at both sides are written as: The pixel points on the image are discontinuous, in order to measure the degree of symmetry, define: wherein: is the difference between the gray value of the point and the average gray value of the surrounding points the square of the difference between the gray value of the point and the average gray value of the surrounding points, within a certain range the degree of symmetry is expressed as: In the formula, T represents the number of pixel points calculated, S is advanced by one pixel point each time, and the number of pixel points calculated is set by setting the value of T; the symmetry of the normal direction gray value of the candidate point is evaluated according to the size of S, the smaller S is, the higher the symmetry is, and the higher the score is, with a full score of 100 points. In the formula, T represents the number of pixel points calculated, S is advanced by one pixel point each time, and the number of pixel points calculated is set by setting the value of T; the symmetry of the normal direction gray value of the candidate point is evaluated according to the size of S, the smaller S is, the higher the symmetry is, and the higher the score is, with a full score of 100 points. The gray value in the image is mapped into a gray score value ranging from 0 to 100; wherein the score value of the gray value less than 25 is 100, the gray value of the pixel point 25-125 is linearly mapped into the score value from 100 to 0, and the score of the pixel point higher than 125 is 0; After calculating the score of each point according to the above method, the score is adjusted, and the specific method is: The highest score of the first ranking is 100, and the score decreases according to the ranking, and the score of the last ranking is 0, each item is adjusted and then weighted and averaged, and finally the final score is adjusted again according to the above method; the weighted average score is divided by 100 to obtain the confidence of the candidate point. In the step S3, first, threshold limit is performed, and the points with confidence (0-1) lower than a certain threshold value are removed, and non-maximum suppression is performed on the remaining points; the non-maximum suppression includes two aspects, which are feature vector direction non-maximum suppression and confidence non-maximum suppression; 3. The method of claim 1, wherein, Second, the step of confidence non-maximum suppression is: by selecting the candidate points with higher score in the neighborhood candidate points of a certain candidate point, and discarding the candidate points with lower score; the candidate points most likely to be crack center points are obtained, so as to sparsify the candidate points, reduce the candidate points useless for crack connection, and improve the connection accuracy of the candidate points. First, the step of non-maximum suppression of the eigenvector is: first, the direction around a point is divided into 6 regions, and the points within a certain range of the point are divided into 6 categories according to the direction of the eigenvector pointed by the direction, and the points are divided into 6 categories; the points in each region are one category, and within a certain range of the point, only the point with the eigenvector pointing to the most region is reserved by using the non-maximum suppression NMS method. In the step S4, the step of connecting the candidate points is:

4. The method of claim 1, wherein, In the step S5, the pseudo crack removal method is: Sort the candidate points by confidence from large to small, set search step and range along the tangent direction of the first point and search the candidate points in the opposite direction; if the next candidate point is not the first and second points of the straight line, calculate the angle formed by the three-point line, if the angle is greater than a certain threshold and the difference between the gray value of the next candidate point and the gray value of the previous point is less than a certain threshold , then the next candidate point and the previous candidate point are regarded as the center points of the same crack, and are connected; repeat the above steps to connect each candidate point on each crack point in turn until the search range is less than the threshold or there is no candidate point within a certain range of the last candidate point, then it is regarded as the end of the crack; The steps for connecting fracture cracks are as follows: Starting from the first crack, calculate the distance between the first and last points of other cracks and the first and last points of this crack in turn, and obtain the two points with the shortest distance; First, calculate the angle formed by the extension line of the two points and the line connecting the adjacent point on the straight line; Second, calculate the distance between the two points. , The threshold representing the distance between the endpoints, and the angle threshold formed by two adjacent fracture endpoints and a fracture point adjacent to one of the endpoints; for values ​​less than the threshold... , Connect the ends of the cracks; place the merged cracks at the end, and then start calculating and connecting from the next crack until none of the cracks meet the connection conditions, thus completing the connection of the fracture cracks.

5. The method of claim 1, wherein, ​ In the first In the first crack, let the distance between its two nodes be calculated by the following formula: In the formula: Indicates the first The length of each crack is calculated using the above formula. Compared to the longest crack in the structure, shorter cracks are directly removed. Furthermore, in complex rock mass fracture situations, many unexpected noise points and pseudo-cracks exist. Although previous steps failed to remove them, since they are not true cracks, their pixel grayscale values ​​are much higher than those of true crack locations. These pseudo-cracks are removed by calculating the average grayscale value, where the average grayscale value is calculated as follows: wherein represents the average gray value of the represents the average gray value of the represents the average gray value of the filter out false fissures above the threshold.

Images