CN-121998987-A - Structure detection-oriented crack global optimization reconstruction method and system

CN121998987ACN 121998987 ACN121998987 ACN 121998987ACN-121998987-A

Abstract

The application discloses a method and a system for global optimization reconstruction of cracks in structure detection. The method comprises the steps of preprocessing a structural surface image, detecting cracks, analyzing a communication area to obtain a discrete crack fragment set, extracting geometric features of each fragment, calculating space adjacent distances of the fragments to construct a candidate connection set and a crack connection diagram, constructing feature vectors based on the connection probability features, combining a likelihood model and a structure priori model to form a probability diagram model, carrying out iterative optimization through a maximum posterior probability inference algorithm dynamically updated by a structural state after the energy is converted into an energy minimization problem, screening an optimal connection relation set, and combining the crack fragments to realize global optimization reconstruction. The method and the device can improve the accuracy and the continuity of crack detection in a complex scene.

Inventors

XU XIANG
Huang Sucan
LI YUEFEI
GONG ZHI
LUO WENMIN
WANG XIANGLIN
WANG YANG
LI EN
WANG QIN
Nie Mincan

Assignees

湖南信息学院

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (9)

1. The global optimization reconstruction method for the crack of the structure detection is characterized by comprising the following steps of: S1, preprocessing image data of the surface of a structure to be detected, performing crack detection on the preprocessed image, and analyzing a communication area of a crack detection result to obtain a discrete crack fragment set; S2, extracting geometric features of the crack fragments, namely extracting the geometric features of each crack fragment in the crack fragment set to form geometric feature information of each crack fragment; S3, constructing a candidate connection set, namely calculating a space adjacent distance for each crack segment pair according to the geometric characteristic information of the crack segments, screening out crack segment pairs meeting the conditions according to a preset space neighborhood threshold value, and constructing the candidate connection set; S4, establishing a crack connection diagram, namely taking a crack segment as a node and crack segment pairs in a candidate connection set as edges, and establishing the crack connection diagram; s5, constructing a connection probability feature vector, namely extracting connection probability features of candidate connections corresponding to each edge in the crack connection graph and constructing the connection probability feature vector; S6, constructing a probability map model, namely constructing a crack connection likelihood model based on the connection probability feature vector, constructing a structure prior model by combining the crack physical topological characteristics, and fusing the likelihood model and the prior model to obtain the probability map model of the crack structure; s7, constructing a global energy function, namely converting the maximum posterior probability estimation problem of the probability map model into an energy minimization problem, and constructing a corresponding crack structure global energy function; S8, energy minimization iterative optimization, namely adopting a maximum posterior probability inference algorithm based on dynamic update of the structural state, performing iterative optimization on the crack connection graph with global energy minimization as a target, and screening a connection edge set with minimum global energy; And S9, overall optimization and reconstruction of the crack, namely updating the crack connection diagram according to the optimized connection side set to obtain an optimal crack connection diagram, merging crack fragments according to the connection relation of the diagram, and realizing overall optimization and reconstruction of the crack structure.
2. The method of claim 1, wherein in S1, the preprocessing comprises at least one of image size normalization, gray level normalization, noise filtering and contrast enhancement, wherein the crack detection comprises performing gradient analysis on the preprocessed standardized image through an edge detection operator to obtain a crack edge response graph, and performing connected region analysis on the crack edge response graph to obtain a discrete crack fragment set.
3. The method of claim 1, wherein in S2, the geometric features include a length of a crack segment, a main direction angle and an end point position, the crack segment length is an accumulated path length obtained by sequentially accumulating euclidean distances of adjacent pixels in the crack segment, the main direction angle is calculated by least squares fitting the crack segment to a straight line model, and the end point position is a pair of coordinates of pixels with the largest euclidean distance in a pixel set of the crack segment.
4. The method of claim 3, wherein in S3, for each crack segment pair, the spatial abutment distance is the minimum euclidean distance in all endpoint combinations of two crack segments, calculated as Wherein 、 The number of the crack fragments is given as the serial number, Is a crack segment And (3) with Is provided for the spatial abutment distance of (c), Is a crack segment First, the The coordinates of the individual end points, Respectively crack fragments First, the And setting the preset space neighborhood threshold according to the image resolution, the crack scale or the average crack width.
5. The method according to claim 1, wherein in S5, the connection probability features include a spatial distance feature, a direction difference feature, a gradient difference feature and a gray difference feature, the spatial distance feature is a spatial adjacent distance of two crack segments, the direction difference feature is an absolute value of a main direction angle difference value of the two crack segments, the gradient difference feature is an absolute value of an average gradient intensity difference value of a local statistical region of each of the two crack segments, the gray difference feature is an absolute value of an average gray value difference value of a local statistical region of each of the two crack segments, the local statistical region is a region obtained by performing a specific radius neighborhood expansion with a pixel set of a corresponding crack segment as a center, and the average gradient intensity and the average gray value are both obtained by statistics in the local statistical region.
6. The method of claim 1, wherein in S6, the crack junction likelihood model is established for any crack segment pair And (3) with The connection probability feature vector between the two Each element in the crack is weighted and summed and then converted into numerical values in the interval of 0 to 1 through a monotonic probability mapping function to obtain the crack fragment And (3) with Probability of existence of connection between Wherein Representing crack fragments And (3) with The connection relation between the two is variable when the crack fragments And (3) with When there is a connection relationship between them Otherwise And performing product operation based on the existence probabilities of the connection between all crack segment pairs to obtain a crack connection likelihood model, wherein the formula is as follows: Wherein Is shown in Under the condition of (2) The probability of the occurrence of the presence of a defect, Representing a set of connected probability feature vectors for all crack segment pairs, Representing the set of connection state variables for all crack segment pairs, representing the overall connection state of the crack connection graph, Representing a set of candidate connections; the structure priori model is as follows: Wherein Crack junction graph Is used to determine the prior probability of (c) for a given channel, Representing the number of bifurcation nodes in the crack junction graph, A measure of complexity representing the crack structure, And (3) with The bifurcation weight and the complexity weight respectively, Representing a positive correlation, the structure prior model represents And (3) with In positive correlation, i.e Is of the value of (2) Is increased by an increase in (a); The likelihood model and the prior model are fused based on a Bayesian inference principle, and the posterior probability model of the crack structure is obtained by fusing the likelihood model and the prior model, wherein the formula is as follows: Wherein Representing the set of crack fragments observed Crack connection structure under condition Is used to determine the posterior probability of (1), And taking the posterior probability model as a probability map model of the crack structure.
7. The method of claim 6, wherein in S8, the specific process of the iterative optimization of energy minimization is that initial energy contributions of all candidate connections are calculated according to existence probabilities of the candidate connections, the candidate connections are ranked according to the initial energy contributions from small to large, a crack connection graph with a connection edge set being an empty set is initialized, candidate connections with the smallest energy contributions are sequentially selected from the ranked candidate connections, global energy variation of the crack connection graph after the candidate connections are added is calculated, if the global energy variation is smaller than 0, the candidate connections are accepted and the connection edge set of the crack connection graph is updated, the prior correlation of crack structures and the energy contributions of the candidate connections are recalculated after each update, iteration is continued after the candidate connections are reordered until the candidate connections are traversed or the global energy function converges, and the optimal connection edge set is obtained.
8. The method of claim 7, wherein the initial energy contribution is calculated as: Wherein Is a crack segment And (3) with And the global energy change is the difference value of the global energy function value of the crack connection graph after the candidate connection is added and before the candidate connection is added.
9. A crack global optimization reconstruction system for structure detection, comprising: The image preprocessing and crack fragment extracting module is used for preprocessing image data of the surface of the structure to be detected, performing crack detection on the preprocessed image, and analyzing a communication area of a crack detection result to obtain a discrete crack fragment set; the crack segment geometrical feature extraction module is used for extracting geometrical features for each crack segment in the crack segment set respectively to form geometrical feature information of each crack segment; the candidate connection set construction module is used for calculating the space adjacent distance for each crack segment pair according to the geometric characteristic information of the crack segments, screening out crack segment pairs meeting the conditions according to a preset space neighborhood threshold value and constructing a candidate connection set; The crack connection diagram establishing module is used for establishing a crack connection diagram by taking a crack fragment as a node and a crack fragment pair in the candidate connection set as an edge; The connection probability feature vector construction module is used for extracting connection probability features of candidate connections corresponding to each edge in the crack connection graph and constructing connection probability feature vectors; the probability map model construction module is used for establishing a crack connection likelihood model based on the connection probability feature vector, constructing a structure priori model by combining the crack physical topological characteristics, and fusing the likelihood model and the priori model to obtain a probability map model of the crack structure; the global energy function construction module is used for converting the maximum posterior probability estimation problem of the probability map model into an energy minimization problem and constructing a corresponding crack structure global energy function; The energy minimization iterative optimization module is used for carrying out iterative optimization on the crack connection graph by adopting a maximum posterior probability inference algorithm based on dynamic update of the structural state and taking global energy minimization as a target, and screening a connection edge set which enables the global energy to be minimum; And the crack global optimization reconstruction module is used for updating the crack connection diagram according to the optimized connection side set to obtain an optimal crack connection diagram, merging crack fragments according to the connection relation of the diagram, and realizing global optimization reconstruction of the crack structure.

Description

Structure detection-oriented crack global optimization reconstruction method and system Technical Field The application relates to the technical field of computer vision and structural surface defect detection, in particular to a crack global optimization reconstruction method and system for structural detection. Background Cracks are common surface defects in engineering structures such as concrete structures, bridges, tunnels, industrial components and the like, and the generation of the cracks is generally closely related to factors such as stress state of the structure, material performance, construction quality, long-term environmental effects and the like. The occurrence of cracks is often an important early signal of structural damage or performance degradation, so that the detection and analysis of the cracks on the surface of the structure are timely and accurately carried out, and the method has important significance for guaranteeing the safe operation of the engineering structure. Traditional crack detection mainly relies on a manual inspection mode, and a professional performs visual inspection on the surface of the structure and records the position, length and morphological characteristics of the crack. However, the method is low in efficiency, and the detection result is easily influenced by factors such as experience of detection personnel, detection environment and subjective judgment, so that the detection requirements of high efficiency and high consistency are difficult to meet in a large-scale infrastructure inspection task. With the development of computer vision technology, the automatic detection of structural cracks by using image processing and automatic recognition technology gradually becomes a research hot spot. In the existing crack detection method based on computer vision, a traditional image processing technology is adopted in one type of method, and a crack region is extracted through edge detection, threshold segmentation, morphological operation and other modes. For example, in the process of crack detection, gray abrupt regions in an image are often extracted by an edge detection algorithm to obtain crack edge information, wherein typical methods include Canny edge detection and other algorithms. The method has the advantages of simple realization, higher calculation efficiency and the like, but in the practical application environment, the crack edge detection result is often broken or discontinuous due to the factors such as complex texture, illumination change, noise interference and the like on the surface of the structure, so that the complete extraction of the crack structure is affected. In recent years, with the development of deep learning technology, a part of research is beginning to segment or detect a crack image by using a convolutional neural network so as to improve the accuracy of crack identification. For example, the multi-scale characteristics are extracted through a deep convolutional neural network, so that the crack recognition capability under the complex background condition can be improved to a certain extent. However, since cracks generally exhibit structural features that are elongated, low contrast, and unevenly distributed, fine crack information is easily lost during deep network multi-layer downsampling, resulting in underexpression of continuity of the crack structure. Whether the conventional image processing method or the crack detection method based on deep learning, the result thereof is generally output in the form of a crack edge map or a crack segmentation map. In these results, the crack structure often appears as a plurality of discrete crack segments, and there are often different degrees of fracture between the crack segments due to image noise, occlusion, and edge detection errors. If the crack analysis is performed only by relying on the original edge detection result or the segmentation result, it is difficult to accurately restore the complete crack path, so that the calculation of key parameters such as the crack length, the trend, the structural morphology and the like is affected. For the crack segment fracture problem, the existing method is used for connecting the crack segments through morphological closing operation, region growth or connection strategy based on distance rules. However, most of these methods rely on simple conditions such as local geometric rules, e.g., endpoint distance or direction difference, to make connection decisions, and lack systematic modeling of the overall structural features of the crack. When complex texture background exists on the surface of the structure or a plurality of cracks are close to each other, the problem of incorrect connection or connection omission easily occurs, and therefore the accuracy of crack structure recovery is reduced. In general, the existing crack detection technology still has the following defects that a large number of discrete crack fragments are frequen