Intelligent identification method and system for cracks of engineering structures

Abstract

The application discloses an engineering structure crack intelligent identification method and system, relates to the technical field of engineering structure health monitoring, and discloses the engineering structure crack intelligent identification method and system, realizes intelligent identification of cracks through the following steps: acquiring an engineering structure image, extracting features, initializing a graph structure, iteratively inferring connection probability, constructing a topology graph, optimizing physical constraints and performing semantic analysis, can improve the accuracy and robustness of crack identification, provides detailed crack topology structure and quantitative semantic information, and meets engineering evaluation requirements.

Description

Technical Field

[0001] This application relates to the field of engineering structural health monitoring technology, and in particular to intelligent identification methods and systems for cracks in engineering structures. Background Technology

[0002] In infrastructure safety assessment practices, the automatic identification of cracks in engineering structures plays a decisive role in ensuring the long-term reliability of buildings, bridges, and other engineering facilities. Current mainstream technologies rely on deep learning models to perform pixel-level image segmentation or target detection to achieve preliminary location of crack areas. However, such methods face fundamental shortcomings in real-world engineering scenarios. First, the models exhibit high vulnerability to environmental variables. Specifically, factors such as uneven lighting, inherent differences in material surface texture, and interference from dirt cover, common at engineering sites, significantly distort the identification results, leading to a sharp decline in model performance when migrating between different projects and making it unable to adapt to complex and changing site conditions. Second, existing outputs are limited to binary outlines or coarse bounding boxes of cracks, completely lacking the ability to analyze the inherent topological structure of cracks. For example, they cannot determine the location of branch nodes, vertex connections, or key engineering parameters such as the precise length and width distribution characteristics of crack segments and their expansion direction trends. This lack of quantitative semantic information hinders the scientific assessment of structural damage, the reasonable prediction of remaining life, and the accurate formulation of maintenance decisions.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide an intelligent method and system for identifying cracks in engineering structures, aiming to improve the accuracy and robustness of crack identification.

[0005] To achieve the above objectives, this application proposes an intelligent method for identifying cracks in engineering structures, the method comprising: Obtain the image of the engineering structure to be identified; The engineering structure image is subjected to feature extraction and primitive decoding to obtain node heatmaps, skeleton segment heatmaps and width field maps; Based on the node heatmap and the skeleton segment heatmap, candidate graph initialization processing is performed to obtain an initial vertex set and a candidate pixel set. Based on the initial vertex set, the candidate pixel set, and the width field graph, iterative graph connectivity inference processing is performed to obtain the connection probability between vertices. Based on the connection probabilities between vertices and the initial vertex set, graph structure construction is performed to obtain an initial crack topology graph; Based on the initial crack topology map and the width field map, physical constraint verification and optimization are performed to obtain an optimized crack topology map. Based on the optimized crack topology map, engineering semantic parameters are parsed to obtain quantitative semantic information about the cracks.

[0006] In one embodiment, the step of performing feature extraction and primitive decoding processing on the engineering structure image to obtain node heatmaps, skeleton segment heatmaps, and width field maps includes: The engineering structure image is subjected to physically guided convolutional enhancement processing to obtain directional enhancement features; The directional enhancement features are subjected to multi-branch collaborative decoding processing, including: The first decoding branch performs node position and type prediction processing on the directional enhancement features to obtain the node heatmap; The second decoding branch performs centerline connectivity prediction processing on the directional enhancement features to obtain the skeleton segment heatmap. The third decoding branch performs pixel-level width prediction processing on the directional enhancement features to obtain the width field map; In the multi-branch collaborative decoding process, the first decoding branch, the second decoding branch, and the third decoding branch interact with each other by sharing intermediate features and cross-attention calculation.

[0007] In one embodiment, the step of performing physically guided convolutional enhancement processing on the engineering structure image to obtain directional enhancement features includes: The engineering structure image is convolved using multiple linear detection convolution kernels with different dominant directions to obtain a multi-directional response map; The multi-directional response maps are subjected to non-maximum suppression processing along their respective dominant directions to obtain refined edge responses; Deep features are extracted from the engineering structure image using a pre-trained backbone feature extraction network; The refined edge response is weighted and fused with the depth features to obtain the directional enhancement features.

[0008] In one embodiment, information interaction is performed through cross-attention computation, including: During the decoding process, the intermediate node features generated by the first decoding branch are used as the query vector, and the intermediate skeleton features generated by the second decoding branch are used as the key vector and value vector. Attention calculation is performed to generate the first update feature. The intermediate skeleton feature generated by the second decoding branch is used as the query vector, and the intermediate width feature generated by the third decoding branch is used as the key vector and value vector. Attention calculation is performed to generate the second update feature. The node heatmap is generated based on the first update feature, and the skeleton segment heatmap is generated based on the second update feature.

[0009] In one embodiment, the step of performing iterative graph connectivity inference processing based on the initial vertex set, the candidate pixel set, and the width field graph to obtain the connection probabilities between vertices includes: Based on the position of each vertex in the initial vertex set, the corresponding width value is extracted from the width field map to construct the initial vertex features; Based on the initial features of the vertices and the geometric spatial relationship represented by the candidate pixel set, the initial adjacency relationship between the vertices is constructed. Perform multiple rounds of message passing iterative processing. In each iteration: Each vertex sends a message containing its own characteristics to its neighboring vertices according to the initial adjacency relationship; Each vertex receives the messages from its neighboring vertices and updates its own vertex features based on all the received messages and its own features; Based on the updated vertex features and the skeleton segment heatmap, the association score between each pair of vertices based on feature similarity and path coherence is recalculated to dynamically update the connection probability between vertices. After a preset number of rounds of message passing iterations, the final inter-vertex connection probability is output.

[0010] In one embodiment, the weight of a vertex sending a message to its neighboring vertices is calculated using a preset attention function, the input of which is the features of the sending vertex and the receiving vertex. When a vertex updates its own features, it performs a weighted aggregation of all the received messages, and the weight is the weight calculated by the attention function.

[0011] In one embodiment, the step of performing graph structure construction processing based on the inter-vertex connection probabilities and the initial vertex set to obtain an initial crack topology graph includes: Based on the connection probability between vertices, vertex pairs with a connection probability higher than the connection threshold are selected; For each selected vertex pair, extract the pixel connectivity path connecting the vertex pair from the skeleton segment heatmap, and convert the pixel connectivity path into an ordered point sequence; The vertices connected to each other in the initial vertex set are determined as nodes of the initial crack topology graph, and the ordered sequence of points is determined as edges connecting the corresponding nodes, thereby constructing the initial crack topology graph.

[0012] In one embodiment, the step of performing physical constraint verification and optimization processing based on the initial crack topology map and the width field map to obtain an optimized crack topology map includes: Based on the ordered list of points associated with each edge in the initial crack topology map and the width field map, a synthetic crack rendering map is generated; The synthetic crack rendering map is compared with the skeleton segment heat map and the width field map respectively to obtain the rendering consistency difference; Based on the topology of the initial crack topology graph, the degree of violation of the preset physical constraints is calculated to obtain a physical violation metric. The preset physical constraints include crack direction smoothness constraints, crack width gradient constraints, and network connectivity constraints. Based on the rendering consistency difference and the physical violation metric, a joint optimization objective is constructed; Guided by minimizing the joint optimization objective, the vertex positions, edge connections, and edge widths in the initial crack topology graph are iteratively adjusted until the convergence condition is met, so as to obtain the optimized crack topology graph.

[0013] In one embodiment, based on the optimized crack topology map, engineering semantic parameter parsing processing is performed to obtain quantified semantic information of the cracks, including: Identify the leaf nodes and branch nodes in the optimized crack topology graph; For each leaf node, trace its path to the nearest branch node or another leaf node in the optimized crack topology graph, and identify the path as a crack segment. Calculate the length, average width, and width variation trend along the path for each crack segment; The direction of crack propagation is determined based on the trend of the width along the path; wherein, the end where the width shows a decreasing trend is determined as the direction of the crack propagation tip; wherein, the direction of the crack propagation tip, the length, and the average width are used as the quantitative semantic information of the crack.

[0014] Furthermore, to achieve the above objectives, this application also proposes an intelligent identification system for engineering structure cracks, the intelligent identification system for engineering structure cracks comprising: a memory, a processor, and an intelligent identification program for engineering structure cracks stored in the memory and executable on the processor, the intelligent identification program for engineering structure cracks being configured to implement the steps of the intelligent identification method for engineering structure cracks.

[0015] The intelligent crack identification method and system for engineering structures proposed in this application achieves intelligent crack identification by acquiring images of engineering structures, extracting features, initializing graph structures, iteratively inferring connection probabilities, constructing topological graphs, optimizing physical constraints, and performing semantic parsing. This improves the accuracy and robustness of crack identification, provides detailed crack topological structures and quantitative semantic information, and meets the needs of engineering assessment. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an embodiment of the intelligent crack identification method for engineering structures in this application; Figure 2 This is a structural schematic diagram of an embodiment of the intelligent crack recognition system for engineering structures provided in this application.

[0019] Explanation of icon numbers: 10. Memory; 20. Processor.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] In existing technologies, intelligent crack recognition methods for engineering structures have weak generalization ability when applied across different scenarios. They are easily affected by factors such as lighting, material texture, and surface stains, leading to a decrease in recognition accuracy. At the same time, these methods lack an understanding and output of the engineering semantics of cracks. They can usually only output binary images or bounding boxes of cracks, and cannot directly provide quantitative information that is crucial for structural health assessment, such as the topological connectivity of cracks, precise dimensions (length, width), and morphological features (direction, branch points). This makes it difficult to meet the needs of automated and refined damage diagnosis in actual engineering.

[0024] Based on this, the embodiments of this application provide a method for intelligent identification of cracks in engineering structures, referring to... Figure 1 The intelligent crack identification method for engineering structures includes steps S100 to S700, wherein: Step S100: Obtain the image of the engineering structure to be identified; Step S200: Perform feature extraction and primitive decoding on the engineering structure image to obtain node heatmap, skeleton segment heatmap and width field map; Step S300: Based on the node heatmap and the skeleton segment heatmap, perform candidate map initialization processing to obtain an initial vertex set and a candidate pixel set; Step S400: Based on the initial vertex set, the candidate pixel set, and the width field graph, perform iterative graph connectivity inference processing to obtain the connection probability between vertices; Step S500: Based on the connection probabilities between vertices and the initial vertex set, perform graph structure construction processing to obtain an initial crack topology graph; Step S600: Based on the initial crack topology map and the width field map, perform physical constraint verification and optimization processing to obtain an optimized crack topology map; Step S700: Based on the optimized crack topology map, perform engineering semantic parameter parsing processing to obtain the quantitative semantic information of the cracks.

[0025] In this embodiment, the node heatmap represents the image of the potential node (e.g., endpoint, branch point) locations and their confidence scores in the crack structure. The skeleton segment heatmap represents the image of the connected region of the crack centerline and its confidence score. The width field map represents the predicted crack width at each pixel in the image. The initial vertex set represents the set of key crack points identified after preliminary processing. The candidate pixel set represents the set of pixels that could potentially form crack connection paths. The inter-vertex connection probability is used to measure the likelihood of a crack connection path between two vertices. The initial crack topology graph represents the crack structure graph constructed based on the preliminary connection probability, containing nodes and edges. The optimized crack topology graph represents the crack structure graph after physical constraint verification and adjustment, which better reflects actual physical characteristics. The quantified semantic information of the crack is used to describe engineering parameters such as crack length, width, and propagation direction.

[0026] In this embodiment, the intelligent crack recognition method for engineering structures first acquires an image of the engineering structure to be identified. This image can be raw image data directly acquired through a digital camera, scanner, or other image acquisition device. Subsequently, feature extraction and primitive decoding are performed on the engineering structure image to obtain node heatmaps, skeleton segment heatmaps, and width field maps. Specifically, traditional image processing algorithms, such as edge detection (e.g., Canny, Sobel operators), morphological operations (e.g., skeleton extraction), and simple pixel intensity analysis, can be used to generate preliminary node, skeleton, and width information maps. For example, node heatmaps can be generated using corner detection algorithms, skeleton segment heatmaps can be generated using thinning algorithms, and width field maps can be obtained through distance transformation or local grayscale profile analysis.

[0027] Furthermore, based on the node heatmap and the skeleton segment heatmap, candidate graph initialization processing is performed to obtain an initial vertex set and a candidate pixel set. This processing may include threshold segmentation of the node heatmap, determining pixels above a preset threshold as the initial vertex set. Simultaneously, threshold segmentation of the skeleton segment heatmap is performed, determining pixels within connected regions above a preset threshold as the candidate pixel set.

[0028] Based on this, iterative graph connectivity inference is performed using the initial vertex set, the candidate pixel set, and the width field graph to obtain the connection probabilities between vertices. This process may include calculating the shortest path length between any two vertices in the initial vertex set within the candidate pixel set, and combining this with the average width of pixels along the path in the width field graph to initially estimate the connection probability. Subsequently, simple local optimizations can be performed; for example, when multiple paths exist between two vertices, the shortest or widest path is selected as the connection.

[0029] Therefore, based on the connection probabilities between vertices and the initial vertex set, a graph structure construction process is performed to obtain an initial crack topology graph. This process can set a fixed connection probability threshold, treating all vertex pairs with connection probabilities higher than this threshold as connections. Then, these connected vertex pairs and the straight line segments or simple pixel paths between them are directly used as edges of the initial crack topology graph.

[0030] Furthermore, based on the initial crack topology map and the width field map, physical constraint verification and optimization are performed to obtain an optimized crack topology map. This process may include manually checking the initial crack topology map and judging, based on experience, whether it conforms to the physical characteristics of cracks, such as cracks should not abruptly interrupted or their width should not change drastically. For cases that do not conform, vertex positions or connection relationships can be manually adjusted.

[0031] In this embodiment, the optimized crack topology map is ultimately used for engineering semantic parameter parsing to obtain quantitative semantic information about the cracks. This process may include manually measuring the length and width of each crack segment in the optimized crack topology map. For example, measurement tools in image processing software can be used to manually click on the start and end points of the crack to calculate the length, and the width can be manually measured at different locations on the crack and the average value calculated.

[0032] In this embodiment, multi-stage processing of the engineering structure image is performed, from primitive decoding to graph connectivity reasoning, and then to topological graph construction and physical constraint optimization, ultimately achieving the parsing of engineering semantic parameters of cracks. This overcomes the limitations of traditional methods in cross-scenario generalization and provides quantitative information such as the topological connectivity of cracks, precise dimensions, and morphological features, meeting the needs of automated and refined crack damage diagnosis in practical engineering.

[0033] In one feasible implementation, the step of performing feature extraction and primitive decoding on the engineering structure image to obtain a node heatmap, a skeleton segment heatmap, and a width field map includes: performing physically guided convolutional enhancement processing on the engineering structure image to obtain directional enhancement features; and performing multi-branch collaborative decoding processing on the directional enhancement features, including: a first decoding branch performing node position and type prediction processing on the directional enhancement features to obtain the node heatmap; a second decoding branch performing centerline connectivity prediction processing on the directional enhancement features to obtain the skeleton segment heatmap; and a third decoding branch performing pixel-level width prediction processing on the directional enhancement features to obtain the width field map; wherein, during the multi-branch collaborative decoding process, the first decoding branch, the second decoding branch, and the third decoding branch interact with each other by sharing intermediate features and cross-attention calculation.

[0034] In this embodiment, the engineering structure image undergoes physically guided convolutional enhancement processing to obtain directional enhancement features. This processing aims to extract and enhance directional features related to the crack structure from the original engineering structure image. Cracks, as linear defects, often exhibit specific orientations and continuity in images. Physically guided convolutional enhancement processing means incorporating prior knowledge of the crack's physical morphology into the design or learning process of the convolutional kernel. For example, by using convolutional kernels that respond to edges or lines in specific directions, the geometric characteristics of the crack can be effectively highlighted, while suppressing background noise and irrelevant textures. For instance, a series of predefined filters with different directional sensitivities can be used to process the image to capture the crack's response intensity in different directions.

[0035] In this embodiment, the directional enhancement features are subjected to multi-branch collaborative decoding. This process is an efficient parallel decoding architecture, the core of which lies in simultaneously and collaboratively parsing multiple crack primitive information from a unified directional enhancement feature. By setting up multiple dedicated decoding branches, each branch is responsible for predicting a specific crack attribute, thereby avoiding information loss and error accumulation that may occur in serial processing. This parallel structure helps improve processing efficiency and lays the foundation for subsequent information interaction.

[0036] In this embodiment, the first decoding branch performs node location and type prediction processing on the directional enhancement features to obtain the node heatmap. This branch focuses on identifying key topological points, i.e., nodes, in the crack network. Nodes can be the start point, end point, intersection point, or branch point of a crack. The node heatmap is typically a two-dimensional probability distribution map, where a higher pixel value indicates a greater probability that the location is a crack node. Type prediction further distinguishes the specific topological attributes of these nodes, for example, distinguishing between crack endpoints and intersection points. This branch can be implemented based on a deep learning model, using a series of convolutional layers and upsampling layers to ultimately output a probability map with one or more channels, indicating the presence and type of nodes respectively.

[0037] In this embodiment, the second decoding branch performs centerline connectivity prediction processing on the orientation enhancement features to obtain the skeleton segment heatmap. This branch aims to identify the skeleton or centerline of the crack, which represents the geometric path and connectivity of the crack. The skeleton segment heatmap is a two-dimensional probability map, where high-value pixels indicate a higher probability that the location belongs to the crack centerline. Accurate prediction of skeleton segments provides a foundation for subsequent crack path tracing and topology map construction. This branch can also employ a deep learning decoder structure to map the orientation enhancement features to the probability distribution of skeleton segments.

[0038] In this embodiment, the third decoding branch performs pixel-level width prediction processing on the directional enhancement features to obtain the width field map. This branch is responsible for predicting the width information of the crack at each pixel location. The width field map is a two-dimensional numerical map, where the value of each pixel directly represents the estimated width of the crack at that location. Accurate pixel-level width information is crucial for the quantitative evaluation of cracks. This branch can be implemented as a regression network whose output layer directly predicts continuous width values.

[0039] In this embodiment, it is worth noting that during the multi-branch collaborative decoding process, the first decoding branch, the second decoding branch, and the third decoding branch interact with each other through sharing intermediate features and cross-attention calculation. This is the key mechanism for achieving "collaboration." Sharing intermediate features means that different decoding branches can access and utilize common or complementary information extracted from each other in the early stages of processing, avoiding repeated extraction from scratch. Cross-attention calculation provides a more refined and dynamic way of information interaction, allowing the feature information of one branch to selectively influence and modify the feature representation of another branch. For example, the prediction of a node branch can utilize the connectivity context provided by the skeleton segment branch to enhance the accuracy of node localization, and vice versa. This interaction mechanism ensures the consistency and mutual support between different primitive information, thereby improving the overall decoding accuracy and robustness. For example, attention weights can be calculated and feature fusion can be performed by using the features of one branch as a query and the features of another branch as keys and values.

[0040] In this embodiment, physically guided convolutional enhancement processing of the engineering structure image effectively captures the inherent directional features of cracks, providing high-quality input for subsequent decoding and thus enhancing the perception of crack fine structures. Secondly, multi-branch collaborative decoding allows for the parallel and efficient extraction of three key crack primitives—node location and type, centerline connectivity, and pixel-level width—from the enhanced features, avoiding information loss and error accumulation that may occur with serial processing. More importantly, by sharing intermediate features and cross-attention computation, different decoding branches can perform real-time information interaction and mutual correction. For example, node prediction can reference the connectivity information of skeleton segments, and width prediction can utilize the geometric context of nodes and skeleton segments. This collaborative mechanism effectively solves the information inconsistency and redundancy problems that may result from independent decoding, ensuring that the acquired node heatmap, skeleton segment heatmap, and width field map are highly coordinated semantically and geometrically, thereby improving the accuracy and robustness of crack primitive extraction and laying a solid foundation for subsequent graph connectivity inference and topology graph construction.

[0041] In one feasible implementation, the step of performing physically guided convolutional enhancement processing on the engineering structure image to obtain directional enhancement features includes: performing convolutional calculations on the engineering structure image using multiple linear detection convolutional kernels with different dominant directions to obtain a multi-directional response map; performing non-maximum suppression processing on the multi-directional response map along its respective dominant direction to obtain a thinned edge response; extracting depth features from the engineering structure image using a pre-trained backbone feature extraction network; and performing weighted fusion processing on the thinned edge response and the depth features to obtain the directional enhancement features.

[0042] In this embodiment, multiple linear detection convolutional kernels with different dominant directions are used to perform convolution calculations on the engineering structure image to obtain multi-directional response maps. These linear detection convolutional kernels are designed to have a high response to linear structures in specific directions. For example, they can contain 8 or 16 convolutional kernels, each corresponding to a specific angle (such as 0°, 22.5°, 45°, etc.), thereby comprehensively capturing crack information in different directions in the image. By performing convolution operations on the engineering structure image, a series of multi-directional response maps can be generated, each highlighting linear features in the image consistent with the dominant direction of the corresponding convolutional kernel.

[0043] In this embodiment, after obtaining the multi-directional response map, in order to further accurately locate the centerline of the crack and eliminate redundant responses, this application performs non-maximum suppression processing on the multi-directional response map along its respective dominant direction to obtain a refined edge response. Specifically, for each pixel in the response map of each direction, its response value is compared with the response values ​​of its neighboring pixels along that direction. If the response value of the pixel is not the maximum value in its neighborhood, it is suppressed (e.g., set to zero), thereby ensuring that each linear feature retains only the strongest, most central response in each direction. Through this processing, refined edge responses can be obtained, which have more accurate positioning and a clearer linear structure.

[0044] In this embodiment, to obtain richer and more semantically informative features, this application extracts deep features from engineering structure images through a pre-trained backbone feature extraction network. This backbone feature extraction network can be a deep convolutional neural network pre-trained on a large image dataset (such as ImageNet), such as ResNet, VGG, EfficientNet, or Swin Transformer. These networks can learn hierarchical representations of images, from low-level edge textures to high-level semantic concepts, thereby providing a comprehensive understanding of cracks and their surrounding environment, enhancing the robustness and generalization ability of the features.

[0045] In this embodiment, to fully utilize the precise orientation and location information provided by the refined edge response and the rich semantic and contextual information provided by the deep features, this application performs a weighted fusion process to obtain orientation enhancement features. This fusion can be achieved in various ways. For example, the refined edge response and deep features can be concatenated along the channel dimension and then integrated through a convolutional layer; or a weighted summation method can be used, i.e., "orientation enhancement feature = w1 × refined edge response + w2 × deep features," where "w1" and "w2" are learnable weights or preset coefficients. Through this weighted fusion, orientation enhancement features that contain both precise crack geometry information and rich contextual semantic information can be generated, providing high-quality input for subsequent multi-branch collaborative decoding.

[0046] In this embodiment, through the above-described technical solution, this application can effectively solve the technical problem of extracting directional features crucial for crack identification from images of complex engineering structures. Utilizing multiple linear detection convolutional kernels with different dominant directions, crack information in different directions within the image can be comprehensively captured. Furthermore, through non-maximum suppression processing along the dominant direction, the centerline of the crack is accurately located, obtaining a refined edge response. Simultaneously, a pre-trained backbone feature extraction network extracts depth features, supplementing rich semantic and contextual information. Finally, the refined edge response and depth features are weighted and fused, resulting in directional enhancement features that not only possess accurate crack geometric localization capabilities but also robustness against complex backgrounds and noise. This improves the accuracy and reliability of subsequent node heatmaps, skeleton segment heatmaps, and width field map predictions, laying a solid foundation for intelligent crack identification in engineering structures.

[0047] In one feasible implementation, information interaction is performed through cross-attention computation, including: during the decoding process, using the intermediate node features generated by the first decoding branch as a query vector and the intermediate skeleton features generated by the second decoding branch as key and value vectors, performing attention computation to generate a first updated feature; using the intermediate skeleton features generated by the second decoding branch as a query vector and the intermediate width features generated by the third decoding branch as key and value vectors, performing attention computation to generate a second updated feature; generating the node heatmap based on the first updated feature and generating the skeleton segment heatmap based on the second updated feature.

[0048] In this embodiment, cross-attention computation is used to establish associations and perform information fusion between different modalities or different features. Here, it allows features from one decoding branch (as a query vector) to focus on and extract relevant information from features from another decoding branch (as key and value vectors). For example, the query vector can be a feature vector, and the key and value vectors can be a set of feature vectors. By calculating the similarity between the query vector and all key vectors, attention weights are obtained. These weights are then applied to a weighted sum of the value vectors, resulting in an output feature that incorporates relevant information. This mechanism makes information interaction more targeted and can capture deep dependencies between different features.

[0049] In this embodiment, the intermediate node features generated by the first decoding branch are used as query vectors, and the intermediate skeleton features generated by the second decoding branch are used as key and value vectors. Attention calculation is performed to generate the first update feature. This aims to enable the node features to fully utilize the connectivity and structural information contained in the skeleton features when predicting node positions and types. For example, skeleton features can provide the overall direction and connection relationships of cracks, which is crucial for accurately determining the starting point, ending point, and intersection points of cracks. In this way, the update of node features no longer relies solely on its own local information but incorporates a more macroscopic crack structure context.

[0050] In this embodiment, similarly, the intermediate skeleton features generated by the second decoding branch are used as the query vector, and the intermediate width features generated by the third decoding branch are used as the key and value vectors, respectively. Attention is then calculated to generate the second updated features. The purpose is to allow the skeleton features to reference the crack geometry information provided by the width features when predicting centerline connectivity. Crack width information can help the skeleton branches better determine crack continuity, branch points, and potential break points, because width plays an important role in the physical morphology of cracks. For example, a sudden narrowing or disappearance of the width may indicate a break or termination of the skeleton.

[0051] In this embodiment, a node heatmap is generated based on the first updated features, and a skeleton segment heatmap is generated based on the second updated features. This means that the final node heatmap and skeleton segment heatmap are generated based on features enhanced and fused through a cross-attention mechanism. These updated features contain complementary information from other relevant branches, making the generated heatmaps more accurate and consistent. For example, the generation of the node heatmap takes into account the connectivity of the skeleton, avoiding the generation of isolated nodes or nodes that do not conform to the crack topology; the generation of the skeleton segment heatmap takes into account the crack width variation, making the skeleton connectivity prediction more consistent with the actual crack morphology.

[0052] In this embodiment, through the above technical solution, during the multi-branch collaborative decoding process, the first, second, and third decoding branches can perform efficient and targeted information interaction through cross-attention calculation. This interaction mechanism allows each branch to fully utilize the complementary information provided by other branches when generating its own feature map. For example, node features can reference the connectivity of the skeleton, and skeleton features can reference the width information of the crack. This effectively solves the problem of isolated information between branches, ensuring a high degree of consistency and physical rationality among node heatmaps, skeleton segment heatmaps, and width field maps. This improves the accuracy and robustness of intelligent crack identification in engineering structures, especially when dealing with complex crack morphologies, enabling the generation of crack primitive representations that are more consistent with reality.

[0053] In one feasible implementation, the step of performing iterative graph connectivity reasoning based on the initial vertex set, the candidate pixel set, and the width field map to obtain the connection probability between vertices includes: extracting the corresponding width value from the width field map according to the position of each vertex in the initial vertex set to construct initial vertex features; constructing initial adjacency relationships between vertices based on the initial vertex features and the geometric spatial relationship represented by the candidate pixel set; performing multi-round message passing iterative processing, in each round of iteration: each vertex sends a message containing its own features to its neighboring vertices according to the initial adjacency relationship; each vertex receives the messages from its neighboring vertices and updates its own vertex features according to all the received messages and its own features; recalculating the association score between each pair of vertices based on feature similarity and path coherence according to the updated vertex features and the skeleton segment heatmap to dynamically update the connection probability between vertices; and outputting the final connection probability between vertices after a preset number of rounds of message passing iterative processing.

[0054] In this embodiment, the step of constructing the initial vertex features aims to assign a physically meaningful initial feature to each potential crack vertex. Each vertex in the initial vertex set represents a key point in the crack network, and its location information is known. The width field map provides a predicted value of the crack width at each pixel in the image. By querying the width field map, the crack width value corresponding to each vertex position can be obtained. For example, if the vertex position is (x, y), then the pixel value extracted from (x, y) in the width field map is the initial width feature of that vertex. This width value, as part of the initial vertex features, provides important physical constraint information for subsequent connection inference, so that the vertex features not only contain geometric position information but also incorporate the morphological properties of the crack.

[0055] In this embodiment, the purpose of constructing the initial adjacency relationships between vertices is to establish a preliminary and reasonable assumption about vertex connectivity before iterative graph connectivity inference begins. The construction of the initial adjacency relationships comprehensively considers the physical characteristics of the vertices (such as their width) and their relative positions in the image space. For example, a distance threshold can be set; if the Euclidean distance between two vertices is less than this threshold, they are considered to be potentially connected. Simultaneously, a candidate pixel set can be incorporated, which may contain pixels representing potential crack paths. If there is a path consisting of candidate pixels with similar width characteristics between two vertices, the likelihood of their initial adjacency can be increased. This initial adjacency relationship provides the basic graph structure for subsequent message-passing iterative processing, avoiding indiscriminate message passing between all vertex pairs, thereby improving computational efficiency and the focus of inference.

[0056] In this embodiment, during the multi-round message passing iteration, the process aims to refine the feature representation of each vertex by exchanging information and aggregating features between vertices using the concept of a Graph Neural Network (GNN), thereby enabling it to better capture local and non-local contextual information. Multiple iterations ensure that information is fully propagated throughout the graph structure, allowing each vertex to "perceive" information from its more distant neighbors and make more global connection decisions. In each iteration, each vertex sends a message containing its own features to its neighboring vertices based on the initial adjacency relationship. This step is the information sending stage in the message passing mechanism. At the beginning of each iteration, each vertex generates a message based on its currently updated features. This message can be a direct copy of its own features or the result of encoding its own features through a learnable transformation function (e.g., a small neural network). Then, this message is sent to all neighboring vertices with which it has an initial adjacency relationship. This mechanism allows each vertex to propagate its local information to its direct neighbors, providing input for the feature updates of its neighbors.

[0057] In this embodiment, each vertex receives messages from its neighboring vertices and updates its own vertex features based on all received messages and its own features. This step is the information aggregation and feature update stage in the message passing mechanism. After receiving messages from all its neighboring vertices, each vertex aggregates them. The aggregation function can be summation, averaging, max pooling, or more complex attention mechanisms. The aggregated information is then combined with the vertex's current features and an update function (e.g., a gated recurrent unit (GRU) or a multilayer perceptron (MLP)) is used to generate the vertex's new features for the current iteration. In this way, the feature representation of each vertex gradually incorporates the contextual information of its neighborhood, becoming richer and more discriminative.

[0058] Based on this, the association score between each pair of vertices is recalculated based on feature similarity and path coherence, according to the updated vertex features and the skeleton segment heatmap, to dynamically update the connection probability between vertices. This step is the core of iterative graph connectivity reasoning. It utilizes the more discriminative vertex features updated through multiple rounds of message passing, combined with external evidence provided by the skeleton segment heatmap, to dynamically evaluate the connection probability between any two vertices. Feature similarity can be measured by calculating the cosine similarity or Euclidean distance between the updated feature vectors of two vertices, reflecting their proximity in the feature space. Path coherence can be evaluated using the skeleton segment heatmap. For example, on a potential connection path between two vertices, a higher response strength in the skeleton segment heatmap indicates a greater likelihood that the path is a real crack, thus indicating stronger path coherence. By weighting and combining the feature similarity score and the path coherence score, a comprehensive association score is obtained, which is used as the updated connection probability between vertices. This dynamic update mechanism allows the connection probability to be continuously optimized with iteration, more accurately reflecting the true connection relationship of cracks.

[0059] In this embodiment, after a preset number of rounds of message passing iterations, the final vertex connection probabilities are output. When the preset number of message passing iterations is reached, it means that information has been sufficiently propagated and aggregated in the graph structure, and vertex features have converged to a relatively stable state. At this point, the system outputs the vertex connection probabilities calculated in the last iteration. These final connection probabilities will serve as the basis for subsequent graph structure construction processing, determining which vertex pairs should be connected to form the edges of the cracks. The preset number of rounds is a hyperparameter, typically optimized through experiments or validation sets to achieve a balance between computational efficiency and connection accuracy.

[0060] In this embodiment, based on the acquisition of the initial vertex set, candidate pixel set, and width field map, the method first extracts the corresponding width value from the width field map according to the position of each vertex in the initial vertex set, thereby constructing an initial feature containing physical attributes for each vertex. This allows subsequent connection judgments to be based not only on geometric information but also on the key physical quantity of crack width, enhancing the discriminative power of the features. Next, based on these initial vertex features and geometric spatial relationships, initial adjacency relationships between vertices are constructed, laying the foundation for message passing iterative processing. Through multiple rounds of message passing iterative processing, each vertex can continuously send messages containing its own features to its neighboring vertices and receive and aggregate messages from neighbors to update its own features. This iterative feature learning process enables the feature representation of each vertex to fully integrate its local and non-local contextual information, effectively overcoming the limitation of insufficient local information in complex crack patterns. Furthermore, in each iteration, based on the updated vertex features and skeleton segment heatmap, the association score between each pair of vertices based on feature similarity and path coherence is recalculated, and the connection probability between vertices is dynamically updated. The skeleton segment heatmap provides direct evidence of the crack centerline. Combined with the updated vertex features, this makes the calculation of connection probabilities more accurate and physically grounded. Ultimately, after a preset number of iterations, the output vertex connection probabilities more accurately reflect the true connection relationships of the cracks, improving the accuracy of crack topology map construction. This provides high-quality input for subsequent physical constraint verification and optimization, as well as engineering semantic parameter parsing, thereby enhancing the overall performance and reliability of the intelligent crack identification method for the entire engineering structure.

[0061] In one feasible implementation, the weight of the message sent by the vertex to its neighboring vertices is calculated by a preset attention function, the input of which is the features of the sending vertex and the receiving vertex; when the vertex updates its own features, it performs a weighted aggregation on all the received messages, and the weight is the weight calculated by the attention function.

[0062] In this embodiment, during message passing, to enable each vertex to selectively focus on information sent by its neighboring vertices, this application introduces message weights. These weights quantify the importance or relevance of each neighboring vertex's message. Specifically, these weights are calculated using a "preset attention function." An attention function is a mechanism that dynamically assigns weights based on input features; its core idea is to allow the model to focus on the most relevant parts of the input. Here, the input to the attention function is the "features of the sending and receiving vertices," meaning the attention mechanism considers the respective feature information of the message sender (sending vertex) and the message receiver (receiving vertex), such as their geometric position, crack width, and local texture features. By comparing and analyzing these features, the attention function learns and outputs a scalar value as the message weight, which reflects the degree to which the sending vertex's information contributes to the receiving vertex updating its own features. For example, if two vertices are spatially close and have similar features, the attention function may assign a higher weight to the messages passed between them. The attention function can take many forms, such as dot-product-based attention mechanisms, additive attention mechanisms, or multi-head attention mechanisms. The specific implementation method can be selected and optimized according to actual needs and computing resources.

[0063] In this embodiment, when a vertex receives messages from multiple neighboring vertices, a "weighted aggregation" approach is used to effectively integrate this information and update its own features. Weighted aggregation refers to linearly combining or nonlinearly transforming all received messages according to their respective weights, rather than simply summing or averaging. The weights here are those calculated by the attention function, ensuring that the importance weights calculated during the message sending phase are consistently applied during the message receiving and aggregation phases. This means that messages deemed more important by the attention function will have a greater influence on vertex feature updates, while less important messages are assigned lower weights, thus reducing their interference with feature updates. This weighted aggregation method enables vertices to more intelligently filter and integrate information from neighbors when updating their own features, thereby obtaining more accurate and robust feature representations. For example, feature updates can be completed by multiplying each received message vector by its corresponding attention weight, summing all weighted message vectors, and then combining this with the vertex's own features for a nonlinear transformation.

[0064] In this embodiment, an attention mechanism is introduced into the iterative graph connectivity inference process using the above-described technical solution, making message passing and feature updating more targeted and effective. Specifically, by calculating message weights using a preset attention function, the importance of messages can be dynamically evaluated based on the features of the sending and receiving vertices, thus avoiding the indiscriminate transmission and processing of all neighbor information and effectively suppressing interference from noise and redundant information. Simultaneously, when a vertex updates its own features, the weights calculated using the attention function are used for weighted aggregation, ensuring that important information dominates the feature fusion process. This allows each vertex to more accurately integrate key information from its neighbors, resulting in a more discriminative and robust feature representation. This refined message passing and feature aggregation mechanism improves the efficiency and accuracy of iterative graph connectivity inference, making the final vertex connection probabilities more reliable. This lays a solid foundation for subsequent graph structure construction and crack topology optimization, ultimately improving the overall accuracy and robustness of intelligent crack identification in engineering structures.

[0065] In one feasible implementation, the step of performing graph structure construction processing based on the vertex connection probability and the initial vertex set to obtain an initial crack topology graph includes: filtering vertex pairs with connection probabilities higher than a connection threshold according to the vertex connection probability; for each filtered vertex pair, extracting the pixel connected path connecting the vertex pair in the skeleton segment heatmap, and converting the pixel connected path into an ordered list of points; determining the vertices connected to each other in the initial vertex set as nodes of the initial crack topology graph, and determining the ordered list of points as edges connecting the corresponding nodes, thereby constructing the initial crack topology graph.

[0066] In this embodiment, vertex pairs with connection probabilities higher than a connection threshold are selected based on the connection probabilities between vertices. This step aims to identify connections with sufficiently high confidence from all possible vertex pairs, thereby filtering out noisy or uncertain connections and laying the foundation for subsequent graph structure construction. Specifically, a predefined connection threshold can be set, such as 0.5, 0.7, or 0.9, which can be adjusted according to the actual application scenario and the desired recognition accuracy. For any two vertices Vi and Vj, if the connection probability P(Vi, Vj) between them is greater than this threshold, then this pair of vertices is considered a potential connection pair. This threshold can also be optimized and determined during the training phase using machine learning methods, such as cross-validation or grid search.

[0067] In this embodiment, for each selected vertex pair, the pixel connectivity path connecting the vertex pair is extracted from the skeleton segment heatmap, and the pixel connectivity path is converted into an ordered list of points. The purpose of this step is to find the actual, pixel-level connectivity path in the skeleton information of the image after identifying the potentially connected vertex pairs, and to structure it into a data form that can be used for graph construction. The skeleton segment heatmap provides connectivity information for the crack centerline. For each selected vertex pair, a graph search algorithm (such as Dijkstra's algorithm or A* algorithm) can be used to find the shortest or most reliable pixel path from one vertex to another on the skeleton segment heatmap. During the search process, the pixel values ​​of the skeleton segment heatmap can be used as the weights or connectivity indicators of the path. Once a path is found, all pixels on the path are arranged in order of their position on the path, forming an ordered list of points. For example, if the path passes through pixels (x1, y1), (x2, y2), ..., (xn, yn), then the ordered list of points is [(x1, y1), (x2, y2), ..., (xn, yn)].

[0068] In this embodiment, the vertices interconnected in the initial vertex set are determined as nodes of the initial crack topology graph, and the ordered point sequence is determined as edges connecting the corresponding nodes, thereby constructing the initial crack topology graph. This formally organizes the connection relationships and path information identified in the previous steps into a graph structure, namely the initial crack topology graph. The initial crack topology graph is a mathematical structure composed of nodes and edges. Nodes correspond to the feature points of the crack (such as intersections and endpoints), and these nodes are interconnected vertices selected from the initial vertex set. Edges represent crack segments; each edge connects two nodes, and an ordered point sequence specifically describes the geometric shape of this crack segment in the image. For example, if vertices A and B are selected as a connection pair, and an ordered point sequence L_AB connecting them is found in the skeleton segment heatmap, then in the initial crack topology graph, A and B will become two nodes, and L_AB will become an edge connecting A and B. This graph structure can be stored using data structures such as adjacency matrices, adjacency lists, or edge lists.

[0069] In this embodiment, through the above technical solution, this application can effectively transform the vertex connection probabilities obtained from iterative graph connection reasoning into an initial crack topology graph with a clear topological structure. Specifically, by setting a connection threshold to filter out high-confidence vertex connections, the introduction of low-confidence potential connections into the crack topology graph is avoided, thereby improving the accuracy of the graph structure. Simultaneously, pixel connected paths are extracted using skeleton segment heatmaps and converted into ordered point sequences as edges, allowing the geometric shape of the crack to be accurately represented in the topology graph, rather than merely as abstract connection relationships. This method systematically organizes discrete vertex and probabilistic connection information into a quantifiable and analyzable graph structure, laying a solid foundation for subsequent physical constraint verification and optimization, as well as engineering semantic parameter parsing, thus improving the structured and accurate nature of crack identification.

[0070] In one feasible implementation, the steps of performing physical constraint verification and optimization based on the initial crack topology map and the width field map to obtain an optimized crack topology map include: generating a synthetic crack rendering map based on the ordered list of points associated with each edge in the initial crack topology map and the width field map; performing difference comparison calculations between the synthetic crack rendering map and the skeleton segment heatmap and the width field map respectively to obtain rendering consistency differences; calculating the degree of violation of preset physical constraints based on the topological structure of the initial crack topology map to obtain a physical violation metric, wherein the preset physical constraints include crack path smoothness constraints, crack width gradient constraints, and network connectivity constraints; constructing a joint optimization objective based on the rendering consistency differences and the physical violation metric; and iteratively adjusting the vertex positions, edge connection relationships, and edge associated widths in the initial crack topology map with the goal of minimizing the joint optimization objective, until the convergence condition is met to obtain the optimized crack topology map.

[0071] In this embodiment, a synthetic crack rendering is generated based on the ordered point sequence associated with each edge in the initial crack topology map and the width field map. The synthetic crack rendering is a simulated crack image generated using computer graphics methods based on the geometric information of the initial crack topology map and the width information of the width field map. Specifically, for each edge in the initial crack topology map, it is associated with an ordered point sequence that defines the centerline path of the crack. By combining the width values ​​of the width field map at these points, a crack region with corresponding width and shape can be simulated. For example, a circle or ellipse can be drawn along each point of the ordered point sequence, using the width value of that point as the diameter, and these shapes can be connected to form a synthetic crack image with a realistic appearance. This rendering visually represents the crack morphology represented by the current topology map, providing a visual basis for subsequent physical constraint verification.

[0072] In this embodiment, the synthesized crack rendering image is compared with the skeleton segment heatmap and the width field map to obtain rendering consistency differences. Rendering consistency differences are an indicator of the degree of matching between the synthesized crack rendering image and the original image features (skeleton segment heatmap and width field map). The skeleton segment heatmap reflects the connectivity of the crack centerline, while the width field map provides pixel-level crack width information. By calculating pixel-level overlap or similarity between the synthesized crack rendering image and the skeleton segment heatmap, it is possible to assess whether the centerline of the synthesized crack is consistent with the detected skeleton segment. Simultaneously, comparing the width information of the synthesized crack rendering image (e.g., extracted from the rendering image through distance transformation or morphological operations) with the original width field map is possible to assess whether the width distribution of the synthesized crack is consistent with the predicted width information. These difference comparison results collectively reflect the degree of geometric and width information agreement between the current crack topology map and the original image features.

[0073] In this embodiment, based on the topological structure of the initial crack topology graph, the degree of violation of preset physical constraints is calculated to obtain a physical violation metric. The physical violation metric quantifies the extent to which the initial crack topology graph does not conform to the inherent physical laws of cracks. The preset physical constraints include crack path smoothness constraints, crack width gradient constraints, and network connectivity constraints. Crack path smoothness constraints can be measured by calculating the angle change or curvature between adjacent line segments on the crack path, avoiding unnatural abrupt turns. Crack width gradient constraints are evaluated by analyzing the rate of change of width values ​​on the crack path, ensuring gradual width changes. Network connectivity constraints are checked using graph theory methods to determine if there are dangling edges, isolated nodes, or connection patterns that do not conform to actual physical meaning in the topology graph. For each constraint, a penalty function can be defined. When the topology graph violates a constraint, the function outputs a positive value, indicating the degree of violation; otherwise, it outputs a zero or near-zero value. The penalty values ​​of all constraints can be weighted and summed to obtain the total physical violation metric.

[0074] Furthermore, a joint optimization objective is constructed based on the rendering consistency difference and the physical violation metric. The joint optimization objective is a comprehensive objective function that combines the rendering consistency difference and the physical violation metric, aiming to simultaneously optimize two aspects of the crack topology map: its matching degree with the original image features and its degree of conformity to physical laws. Typically, this objective function can be expressed as a weighted sum of the rendering consistency difference and the physical violation metric, where the weight coefficients are used to balance the contributions of the two components.

[0075] In this embodiment, finally, guided by minimizing the joint optimization objective, the vertex positions, edge connections, and edge widths in the initial crack topology graph are iteratively adjusted until the convergence condition is met to obtain the optimized crack topology graph. This is an iterative optimization process. The optimization algorithm (e.g., gradient descent, genetic algorithm, or simulated annealing) continuously adjusts the parameters of the initial crack topology graph, including vertex positions, edge connections, and edge widths. In each iteration, the algorithm calculates the value of the joint optimization objective based on the current topology graph and updates the parameters of the topology graph according to the gradient of the objective function or heuristic rules to reduce the objective function value. This process continues until a preset convergence condition is met, such as the change in the objective function value being less than a certain threshold, or reaching the maximum number of iterations.

[0076] In this embodiment, by generating a synthetic crack rendering map and comparing it with the original skeleton segment heatmap and width field map, the degree of consistency between the current topology map and the original image features can be quantified. Simultaneously, by introducing physical constraints such as crack path smoothness constraints, crack width gradient constraints, and network connectivity constraints, the topology map can be evaluated and corrected from a physical rationality perspective. By combining rendering consistency differences with physical violation metrics, a joint optimization objective is constructed. Through iterative adjustments to vertex positions, edge connectivity, and edge association widths, the optimized crack topology map not only more accurately reflects crack information in the image but also becomes more reasonable and reliable in terms of geometric shape and physical properties. This optimization process improves the accuracy and physical rationality of the crack topology map, providing a more reliable foundation for subsequent engineering semantic parameter analysis, thereby improving the accuracy of crack quantification semantic information.

[0077] In one feasible implementation, based on the optimized crack topology map, engineering semantic parameter parsing processing is performed to obtain the quantitative semantic information of the cracks, including: identifying leaf nodes and branch nodes in the optimized crack topology map; for each leaf node, tracing its path to the nearest branch node or another leaf node in the optimized crack topology map, and identifying the path as a crack segment; calculating the length, average width, and width variation trend along the path of each crack segment; determining the crack propagation direction based on the width variation trend along the path; wherein, the end with a decreasing width trend is determined as the crack propagation tip direction; wherein, the crack propagation tip direction, the length, and the average width are used as the quantitative semantic information of the crack.

[0078] In this embodiment, leaf nodes and branch nodes are identified in the optimized crack topology graph. In graph theory, a leaf node is a node with a degree of 1, i.e., a node connected to only one edge, while a branch node is a node with a degree greater than 1, i.e., a node connected to multiple edges. By traversing all nodes in the optimized crack topology graph and calculating the number of connecting edges, these different types of nodes can be accurately distinguished. Identifying leaf nodes and branch nodes is fundamental to understanding the complex topology of cracks. They define the start and end points of cracks, as well as the key locations where cracks intersect or fork, providing clear boundaries for subsequently decomposing the overall crack structure into independently analyzable crack segments.

[0079] In this embodiment, for each leaf node, the path from its nearest branch node or another leaf node is traced in the optimized crack topology graph, and this path is identified as a crack segment. This tracing process can be implemented using graph traversal algorithms such as Depth-First Search (DFS) or Breadth-First Search (BFS), starting from the leaf node and exploring step by step along the connecting edges until the first branch node or another leaf node is encountered. All nodes and edges on this path together constitute an independent crack segment. To ensure the uniqueness of the path and avoid duplication, identified nodes or edges can be marked during the search process. In this way, the complex crack network is effectively decomposed into a series of logically independent and manageable crack segments, thus laying the foundation for subsequent refined quantitative analysis.

[0080] Based on this, this application calculates the length, average width, and width variation trend along the path of each crack segment. The length of a crack segment can be accurately obtained by accumulating the Euclidean distances between adjacent pixels in the ordered sequence constituting the crack segment. The average width is obtained by summing and averaging the width values ​​of all pixels contained in the crack segment in the width field map. The width variation trend along the path can be determined by analyzing the distribution of width values ​​at various points on the crack segment, for example, by calculating the average rate of change of width values ​​from one end of the crack segment to the other, or by performing linear regression analysis to determine whether the width is increasing, decreasing, or relatively stable.

[0081] Furthermore, the direction of crack propagation is determined based on the trend of width variation along the path. The end where the width decreases is identified as the direction of the crack's propagation tip. This determination is based on the physical characteristic that the crack tip is typically the region of minimum width and stress concentration during propagation. By identifying the decreasing width trend, the potential propagation direction of the crack can be accurately indicated, which is of great significance for predicting the future development trend of cracks and assessing structural safety.

[0082] In this embodiment, the direction of the crack's propagation tip, its length, and its average width are ultimately output as the crack's quantitative semantic information. Through the above technical solution, this application transforms the abstract crack topology map into specific, measurable engineering semantic parameters. This quantitative semantic information directly meets the needs of crack damage assessment in engineering practice, enabling engineers to judge the severity, development trend, and potential risks of cracks based on precise quantitative data, rather than just visual observation. This greatly improves the objectivity, accuracy, and efficiency of crack assessment, providing solid data support for subsequent maintenance and reinforcement decisions. Combined with the previously obtained optimized crack topology map, this solution can deeply explore the physical characteristics of cracks at the structural level, enabling crack identification to go beyond mere "discovery" and reach the levels of "understanding" and "quantification," thereby providing comprehensive and in-depth technical support for the health monitoring and maintenance of engineering structures.

[0083] In the embodiments of this application, the intelligent crack identification method for engineering structures achieves intelligent crack identification by acquiring images of engineering structures, extracting features, initializing graph structures, iteratively inferring connection probabilities, constructing topological graphs, optimizing physical constraints, and performing semantic parsing. This method can improve the accuracy and robustness of crack identification, provide detailed crack topological structures and quantitative semantic information, and meet the needs of engineering evaluation.

[0084] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the intelligent identification method for cracks in engineering structures in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0085] This application also provides an intelligent crack recognition system for engineering structures, referencing... Figure 2 The intelligent identification system for engineering structure cracks includes: a memory 10, a processor 20, and an intelligent identification program for engineering structure cracks stored on the memory 10 and executable on the processor 20. The intelligent identification program for engineering structure cracks is configured to implement the steps of the intelligent identification method for engineering structure cracks.

[0086] The intelligent crack identification system for engineering structures provided in this application, employing the intelligent crack identification method for engineering structures in the above embodiments, can improve the accuracy and robustness of crack identification. Compared with the prior art, the beneficial effects of the intelligent crack identification system for engineering structures provided in this application are the same as those of the intelligent crack identification method for engineering structures provided in the above embodiments, and other technical features of the intelligent crack identification system for engineering structures are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0087] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A method for intelligent identification of cracks in engineering structures, characterized in that, The method includes: Obtain the image of the engineering structure to be identified; The engineering structure image is subjected to feature extraction and primitive decoding to obtain node heatmaps, skeleton segment heatmaps and width field maps; Based on the node heatmap and the skeleton segment heatmap, candidate graph initialization processing is performed to obtain an initial vertex set and a candidate pixel set. Based on the initial vertex set, the candidate pixel set, and the width field graph, iterative graph connectivity inference processing is performed to obtain the connection probabilities between vertices. Based on the connection probabilities between vertices and the initial vertex set, graph structure construction is performed to obtain an initial crack topology graph; Based on the initial crack topology map and the width field map, physical constraint verification and optimization are performed to obtain an optimized crack topology map. Based on the optimized crack topology map, engineering semantic parameters are parsed to obtain quantitative semantic information about the cracks.

2. The intelligent crack identification method for engineering structures as described in claim 1, characterized in that, The steps of performing feature extraction and primitive decoding on the engineering structure image to obtain node heatmaps, skeleton segment heatmaps, and width field maps include: The engineering structure image is subjected to physically guided convolutional enhancement processing to obtain directional enhancement features; The directional enhancement features are subjected to multi-branch collaborative decoding processing, including: The first decoding branch performs node position and type prediction processing on the directional enhancement features to obtain the node heatmap; The second decoding branch performs centerline connectivity prediction processing on the directional enhancement features to obtain the skeleton segment heatmap. The third decoding branch performs pixel-level width prediction processing on the directional enhancement features to obtain the width field map; In the multi-branch collaborative decoding process, the first decoding branch, the second decoding branch, and the third decoding branch interact with each other by sharing intermediate features and cross-attention calculation.

3. The intelligent crack identification method for engineering structures as described in claim 2, characterized in that, The steps of performing physically guided convolutional enhancement processing on the engineering structure image to obtain directional enhancement features include: The engineering structure image is convolved using multiple linear detection convolution kernels with different dominant directions to obtain a multi-directional response map; The multi-directional response maps are subjected to non-maximum suppression processing along their respective dominant directions to obtain refined edge responses; Deep features are extracted from the engineering structure image using a pre-trained backbone feature extraction network; The refined edge response is weighted and fused with the depth features to obtain the directional enhancement features.

4. The intelligent crack identification method for engineering structures as described in claim 2, characterized in that, Information interaction through cross-attention computation includes: During the decoding process, the intermediate node features generated by the first decoding branch are used as the query vector, and the intermediate skeleton features generated by the second decoding branch are used as the key vector and value vector. Attention calculation is performed to generate the first update feature. The intermediate skeleton feature generated by the second decoding branch is used as the query vector, and the intermediate width feature generated by the third decoding branch is used as the key vector and value vector. Attention calculation is performed to generate the second update feature. The node heatmap is generated based on the first update feature, and the skeleton segment heatmap is generated based on the second update feature.

5. The intelligent crack identification method for engineering structures as described in claim 1, characterized in that, The steps for performing iterative graph connectivity inference processing based on the initial vertex set, the candidate pixel set, and the width field graph to obtain the connection probabilities between vertices include: Based on the position of each vertex in the initial vertex set, the corresponding width value is extracted from the width field map to construct the initial vertex features; Based on the initial features of the vertices and the geometric spatial relationship represented by the candidate pixel set, the initial adjacency relationship between the vertices is constructed. Perform multiple rounds of message passing iterative processing. In each iteration: Each vertex sends a message containing its own characteristics to its neighboring vertices according to the initial adjacency relationship; Each vertex receives the messages from its neighboring vertices and updates its own vertex features based on all the received messages and its own features; Based on the updated vertex features and the skeleton segment heatmap, the association score between each pair of vertices based on feature similarity and path coherence is recalculated to dynamically update the connection probability between vertices. After a preset number of rounds of message passing iterations, the final inter-vertex connection probability is output.

6. The intelligent crack identification method for engineering structures as described in claim 5, characterized in that, The weight of a vertex sending messages to its neighboring vertices is calculated using a preset attention function, the input of which is the features of the sending and receiving vertices. When a vertex updates its own features, it performs a weighted aggregation of all the received messages, and the weight is the weight calculated by the attention function.

7. The intelligent crack identification method for engineering structures as described in claim 1, characterized in that, The steps for performing graph structure construction processing to obtain an initial crack topology graph based on the vertex connection probabilities and the initial vertex set include: Based on the connection probability between vertices, vertex pairs with a connection probability higher than the connection threshold are selected; For each selected vertex pair, extract the pixel connectivity path connecting the vertex pair from the skeleton segment heatmap, and convert the pixel connectivity path into an ordered point sequence; The vertices connected to each other in the initial vertex set are determined as nodes of the initial crack topology graph, and the ordered sequence of points is determined as edges connecting the corresponding nodes, thereby constructing the initial crack topology graph.

8. The intelligent crack identification method for engineering structures as described in claim 7, characterized in that, The steps for performing physical constraint verification and optimization based on the initial crack topology map and the width field map to obtain the optimized crack topology map include: Based on the ordered list of points associated with each edge in the initial crack topology map and the width field map, a synthetic crack rendering map is generated; The synthetic crack rendering map is compared with the skeleton segment heat map and the width field map respectively to obtain the rendering consistency difference; Based on the topology of the initial crack topology graph, the degree of violation of the preset physical constraints is calculated to obtain a physical violation metric. The preset physical constraints include crack direction smoothness constraints, crack width gradient constraints, and network connectivity constraints. Based on the rendering consistency difference and the physical violation metric, a joint optimization objective is constructed; Guided by minimizing the joint optimization objective, the vertex positions, edge connections, and edge widths in the initial crack topology graph are iteratively adjusted until the convergence condition is met, so as to obtain the optimized crack topology graph.

9. The intelligent crack identification method for engineering structures as described in claim 1, characterized in that, Based on the optimized crack topology map, engineering semantic parameter parsing is performed to obtain the quantitative semantic information of the cracks, including: Identify the leaf nodes and branch nodes in the optimized crack topology graph; For each leaf node, trace its path to the nearest branch node or another leaf node in the optimized crack topology graph, and identify the path as a crack segment. Calculate the length, average width, and width variation trend along the path for each crack segment; The direction of crack propagation is determined based on the trend of the width along the path; wherein, the end where the width shows a decreasing trend is determined as the direction of the crack propagation tip; wherein, the direction of the crack propagation tip, the length, and the average width are used as the quantitative semantic information of the crack.

10. An intelligent crack identification system for engineering structures, characterized in that, The intelligent identification system for engineering structure cracks includes: a memory, a processor, and an intelligent identification program for engineering structure cracks stored in the memory and executable on the processor, wherein the intelligent identification program for engineering structure cracks is configured to implement the steps of the intelligent identification method for engineering structure cracks as described in any one of claims 1 to 9.

Images