A bridge crack intelligent identification method and system based on key points
By improving the PolyHRNet network model and key point clustering algorithm, the accuracy and efficiency problems of bridge crack identification methods in complex scenarios are solved, achieving lightweight and high-precision crack detection that meets real-time requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2025-04-27
- Publication Date
- 2026-06-12
AI Technical Summary
Existing bridge crack identification methods struggle to accurately identify cracks in complex scenarios, and traditional methods are limited by human intervention and parameter dependence, leading to distorted detection results or low efficiency.
A key-point-based intelligent recognition method is adopted, which uses an improved PolyHRNet network model to detect key points of cracks. Combined with data augmentation, loss function optimization and key point clustering algorithm, crack heat map is generated and key points are connected through post-processing to generate crack features.
It improves the accuracy and efficiency of crack detection in complex environments, achieves lightweight and high-precision crack detection, meets real-time requirements, and reduces reliance on manual intervention and parameter adjustment.
Smart Images
Figure CN120451095B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of concrete bridge structure inspection technology, specifically to a method and system for intelligent identification of bridge cracks based on key points. Background Technology
[0002] During construction and service life, concrete structures may develop cracks due to multiple factors, including construction defects, environmental effects, material degradation, and exceeding mechanical limits. Cracks can weaken load-bearing capacity, especially through cracks or network cracks, which may indicate deeper problems such as structural deformation or foundation settlement. Cracks also act as channels for corrosive media such as water and chloride ions to penetrate, accelerating steel reinforcement corrosion, leading to the spalling of the concrete cover, and further reducing the structure's lifespan. Therefore, concrete cracks are not only a direct manifestation of structural pathology but also a core monitoring target for life-cycle management, requiring regular inspection and evaluation. However, traditional detection methods are limited by technical means and manual operation, are highly subjective, and prone to missed or false detections, exhibiting significant shortcomings in efficiency, accuracy, and safety. With the development of intelligent technologies, digital detection systems integrating multi-source data are gradually replacing traditional methods, providing more reliable support for the life-cycle management of concrete structures.
[0003] With the rapid development of new technologies such as drones and artificial intelligence, bridge inspection is moving towards digitalization and intelligence. Utilizing computer vision technology to intelligently identify bridge photographs, locate cracks, and extract information offers advantages over traditional manual inspection, including significantly reducing manpower and improving accuracy and objectivity.
[0004] For example, the Chinese patent "A Bridge Crack Detection Method Based on Image Overlay and Crack Information Fusion" (Patent Application No.: CN 202010886335.3, Publication No.: CN112053331A). This patent describes a method for rapid and high-precision bridge crack detection by combining image overlay and crack information fusion technology with gradient calculation, co-occurring edge extraction, and seed point growth algorithms. This avoids the inefficiency and information loss problems caused by feature point calculation and sampling in traditional methods.
[0005] However, this patent relies on preset parameters such as gradient threshold, co-occurring edge point spacing, and skeleton length to extract crack information. If the crack features in the actual bridge image exceed the preset parameter range, problems such as seed point misjudgment, skeleton breakage, or incorrect connection will lead to distorted detection results. Furthermore, parameter adjustment requires manual intervention, making it difficult to adapt to complex and ever-changing engineering scenarios.
[0006] For example, the Chinese patent "Method and System for Extracting Single-Pixel Edge Curves and Measuring Crack Width from Crack Images" (Patent Application No.: CN202311238469.4, Publication No.: CN117197109A). This patent proposes a method for extracting single-pixel edge curves of cracks based on grayscale thresholding and row-by-row and column-by-column pixel traversal. It generates crack curves by merging row and column edge point sets to improve edge extraction efficiency.
[0007] However, this patented technology extracts crack edge points based on a preset grayscale threshold and a row-by-row, column-by-column pixel traversal rule. This relies on the accuracy of grayscale segmentation and the regularity of the crack shape. If the crack shape deviates from the preset rule or the image quality is insufficient, the grayscale threshold will fail, and the row-by-row, column traversal will not be able to accurately capture the edge points, resulting in decreased extraction efficiency and insufficient robustness. Summary of the Invention
[0008] To address the problems existing in the prior art, this invention provides a bridge crack intelligent identification method and system based on key points, solving the technical problem that existing bridge crack identification methods are difficult to accurately identify in complex scenarios.
[0009] A key-point-based intelligent method for identifying bridge cracks includes:
[0010] Step 1: Collect original crack images to create an original dataset, and then label the original dataset to obtain a crack dataset, which includes crack ground truth maps;
[0011] Step 2: Train the improved PolyHRNet network model using the crack dataset to obtain a crack key point detection model. Input the original crack image into the crack key point detection model and output a crack heat map containing crack key points.
[0012] Step 3: After obtaining the key points of the crack from the crack heat map, the crack identification result is obtained by connecting the key points and generating polylines that reflect the crack characteristics through post-processing operations.
[0013] Further, step 1 includes:
[0014] Step 1.1: Label the images in the original dataset. The original dataset is subjected to horizontal flipping, random scaling, vertical flipping, rotation and other operations to obtain the augmented crack dataset. The labeled multi-line segments are used to generate a crack ground truth map. The generated crack ground truth map is used as the new crack dataset. The crack ground truth map includes information on key points and crack connections, which is used for subsequent network model training.
[0015] Step 1.2: Divide the crack dataset into training and validation sets for training the improved PolyHRNet network model.
[0016] Furthermore, the image processing by the crack key point detection model includes:
[0017] Step 2.1: The original image is first passed through a Conv Block convolutional module with a downsampling factor of 4, and then through a layer1 convolutional module for feature extraction;
[0018] Step 2.2: The feature map output in Step 2.1 continuously introduces new low-resolution branches in the process from stage 1 to stage 4, compresses the spatial dimension through downsampling operation, and adjusts the number of channels through 1x1 convolution;
[0019] Step 2.3: In each stage, the high-resolution branch passes the features to the low-resolution branch through downsampling, and the low-resolution branch passes the semantic information back to the high-resolution branch through upsampling. Stage 3 is repeated four times and stage 4 is repeated three times, finally resulting in four output layers from head1 to head4.
[0020] Step 2.4: After upsampling, head1 to head4 are stacked, then passed through the Neck module, and finally output through the head5 output layer to produce a heat map image containing the key points of the crack.
[0021] Furthermore, the loss function used to train the improved PolyHRNet network model is the MSELoss function, and its calculation formula is shown in the following equation:
[0022]
[0023] In the formula, This represents the predicted value of the i-th sample at pixel position p. Let p represent the true value of the i-th sample at pixel position p, N represent the total number of keypoints, K represent the number of keypoint categories, and p represent the spatial coordinates of the heatmap.
[0024] Furthermore, in the MSELoss loss function, the full loss calculation is retained for the zero-value region of the truth map, and the complete loss calculation is performed for the region of the truth map exceeding 0.8. A loss threshold filtering mechanism of 0.04 is set. The specific calculation formula is as follows:
[0025]
[0026] Furthermore, in step 2.4, the Neck module uses ordinary convolution, and then obtains the predicted heatmap through 4 Basic Block residual modules.
[0027] Further, step 2.1 includes: using max pooling to downsample the ground truth map of the original image size at different ratios to correspond to the output scale of each layer, wherein the pooling kernel size and stride are set as the downsampling factor. While reducing the feature map size, the feature information of the crack key points is preserved. The feature value of each spatial location (i,j) in the feature map Y output after max pooling is calculated as shown in the following formula:
[0028] y ij =max(X (i×s:i×s+k,j×s:j×s+k) )
[0029] In the formula, X represents the input feature map, y ij represents the value of the output feature map at point (i,j), s represents the pooling kernel step size and feature map downsampling rate, and k represents the pooling kernel size.
[0030] Further, step 3 includes:
[0031] Step 3.1: The DBSCAN-based key point clustering algorithm is adopted to effectively identify irregularly shaped clusters and process noise data by analyzing the density distribution of the heat map. The sparse key points generated by the heat map clustering are connected, and the endpoints of the resulting multi-segment lines are matched after the connection is completed.
[0032] Step 3.2: Using the nearest neighbor-first iterative matching algorithm, a global distance matrix is established by iteratively calculating the Euclidean distance of each endpoint combination, and the two closest endpoints are selected in turn to connect them to form a continuous polyline segment;
[0033] Step 3.3: Exclude connected endpoints and iteratively update the distance matrix until the topology reconstruction of all key points is completed;
[0034] Step 3.4: Starting from any endpoint of the polyline, perform vector operation analysis on each pair of adjacent edges. The core detection parameter of the algorithm is defined as the angle θ between the direction vectors of adjacent edges and the angle α between the direction vectors of the interval. Determine whether the angle θ between the direction vectors of adjacent edges or the angle α between the direction vectors of the interval is greater than 90°. If it is greater, it is determined that there is an abnormal connection between the corresponding adjacent edges, and the abnormal connection is adjusted accordingly. Specifically, this includes:
[0035] Step 3.41: Detect the angle between the direction vectors of adjacent sides. When θ > 90° is detected, mark the connection as an abnormal crack topology. At the abnormal connection, prioritize the retention of the edge segment with shorter geometric length, and delete redundant long edges.
[0036] Step 3.42: For a composite connection structure consisting of three consecutive sides, namely sides AB, BC, and CD connected in sequence, calculate the included angle α between the extensions of the first and last sides AB and CD. If α > 90°, the intermediate transition side BC is determined to be a redundant connection, and the side BC is deleted. If α < 90°, the intermediate transition side BC is determined to be a non-redundant connection.
[0037] Furthermore, this includes step 4:
[0038] The dedicated evaluation metric used is CKS (Crack Keypoint Similarity) to evaluate the crack keypoint detection model. In CKS, the F1-score is calculated using precision and recall, and the formula for calculating the F1-score is as follows:
[0039]
[0040] In the formula, F 1_CKS Represented as F1-score, P CKS Represented as spatial accuracy, R CKS Let P be the spatial recall rate. CKS and R CKS The specific calculation method is as follows:
[0041]
[0042] In the formula, N p G(p) represents the number of predicted keypoints. i ) represents the prediction key point p i In the actual heatmap, the activation value N g P(g) represents the number of predicted keypoints. i ) represents the prediction key point p i Activation values in a real heatmap;
[0043] Unlike the normal calculation methods for precision and recall, spatial precision is calculated by looking up the response intensity of the actual labeled points from the predicted keypoint coordinates. The actual value of the point on the actual heatmap corresponding to the predicted keypoint coordinates is used as the spatial precision. Spatial recall is calculated by looking up the response intensity of the predicted points from the predicted heatmap from the actual keypoint coordinates. The predicted value of the point on the predicted heatmap corresponding to the actual keypoint coordinates is used as the spatial recall.
[0044] A bridge crack intelligent recognition system based on key points includes: a dataset creation module, a model training module, a model evaluation module, and a recognition module. The dataset creation module collects original crack images to create an original dataset, and then annotates the original dataset to obtain a crack dataset, which includes a crack ground truth map. The model training module trains a PolyHRNet network model using the crack dataset to obtain a crack key point detection model. The original crack images are input into the crack key point detection model, and a crack heatmap containing crack key points is output. The model evaluation module evaluates the crack key point detection model using a dedicated evaluation metric (CKS). The crack recognition module obtains crack key points from the crack heatmap and then uses post-processing operations to connect the key points and generate polylines reflecting crack characteristics.
[0045] The beneficial effects of this invention include:
[0046] 1. This invention develops a crack key point detection model and designs corresponding modules such as data extraction, data augmentation, and loss calculation for crack key point identification tasks, further improving the model's recognition accuracy when faced with interference such as concrete surface spalling, holes, and marker pen marks.
[0047] 2. The improved PolyHRNet model, as a crack keypoint detection model, has only 9.64M parameters, while the U-net network, commonly used for crack semantic segmentation, has 28M parameters. In terms of CKS evaluation metrics, PolyHRNet achieves an accuracy of 0.81, a recall of 0.94, and a final F1 score of 0.87. This model can achieve 37fps on an RTX 2080s graphics card, while other lightweight segmentation models generally have inference speeds below 10fps on embedded devices. This model successfully achieves the requirements of high accuracy and real-time performance in crack detection while being lightweight.
[0048] 3. This invention uses a crack key point connection optimization method that combines heat map and key point clustering. This method adopts a dual-modal clustering strategy and adds an angle constraint mechanism, which effectively solves the path confusion problem of adjacent or intersecting cracks. By simulating the physical laws of material fracture, the algorithm constraint conditions are designed to achieve a connection method that is more in line with the actual crack propagation law. Attached Figure Description
[0049] Figure 1 This is a flowchart of a bridge crack intelligent identification method based on key points, which is an embodiment of this application.
[0050] Figure 2 This is a schematic diagram of the PolyHRNet network structure involved in the embodiments of this application.
[0051] Figure 3 This is a schematic diagram of the Neck module involved in an embodiment of this application.
[0052] Figure 4 This is an image showing the output of key crack points after the original crack image in the embodiment of this application is processed by the crack key point detection model.
[0053] Figure 5 This is a schematic diagram of the endpoint connection involved in the embodiments of this application.
[0054] Figure 6 Figure 1 shows an example of angle correction involved in the embodiments of this application.
[0055] Figure 7 Figure 2 shows an example of angle correction involved in the embodiments of this application. Detailed Implementation
[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed 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 creative effort are within the scope of protection of this application.
[0057] like Figure 1 As shown, a bridge crack intelligent identification method based on key points includes:
[0058] Step 1: Collect original crack images to create an original dataset, and then label the original dataset to obtain a crack dataset, which includes crack ground truth maps;
[0059] Step 2: Train the improved PolyHRNet network model using the crack dataset to obtain a crack key point detection model. Input the original crack image into the crack key point detection model, and output a crack heatmap containing crack key points. The structure of the PolyHRNet network model is as follows: Figure 2 As shown;
[0060] Step 3: After obtaining the key points of the crack from the crack heat map, the crack identification result is obtained by connecting the key points and generating polylines that reflect the crack characteristics through post-processing operations.
[0061] In another embodiment, step 1 includes:
[0062] Step 1.1: Label the images in the original dataset. The original dataset is subjected to horizontal flipping, random scaling, vertical flipping, rotation and other operations to obtain the augmented crack dataset. The labeled multi-line segments are used to generate a crack ground truth map. The generated crack ground truth map is used as the new crack dataset. The crack ground truth map includes information on key points and crack connections, which is used for subsequent network model training.
[0063] Step 1.2: Divide the crack dataset into training and validation sets for training the improved PolyHRNet network model.
[0064] In another embodiment, the image processing by the crack key point detection model includes:
[0065] Step 2.1: The original image is first passed through a Conv Block convolutional module with a downsampling factor of 4, and then through a layer1 convolutional module for feature extraction;
[0066] Step 2.2: The feature map output in Step 2.1 continuously introduces new low-resolution branches in the process from stage 1 to stage 4, compresses the spatial dimension through downsampling operation, and adjusts the number of channels through 1x1 convolution;
[0067] Step 2.3: In each stage, the high-resolution branch passes the features to the low-resolution branch through downsampling, and the low-resolution branch passes the semantic information back to the high-resolution branch through upsampling. Stage 3 is repeated four times and stage 4 is repeated three times, finally resulting in four output layers from head1 to head4.
[0068] Step 2.4: After upsampling, head1 to head4 are stacked, then passed through the Neck module, and finally output through the head5 output layer to produce a heatmap image containing the key points of the crack. The output effect is as follows. Figure 4 As shown.
[0069] The loss function used to train the improved PolyHRNet network model is the MSELoss function, and its calculation formula is shown in the following equation:
[0070]
[0071] In the formula, This represents the predicted value of the i-th sample at pixel position p. Let p represent the true value of the i-th sample at pixel position p, N represent the total number of keypoints, K represent the number of keypoint categories, and p represent the spatial coordinates of the heatmap.
[0072] In the MSELoss loss function, the full loss calculation is retained for the zero-value region of the truth map, and the complete loss calculation is performed for the region of the truth map with a value greater than 0.8. A loss threshold filtering mechanism of 0.04 is set. The specific calculation formula is as follows:
[0073]
[0074] In step 2.4, the Neck module uses ordinary convolution, and then obtains the predicted heatmap through four Basic Block residual modules. The specific structure of the Neck module is as follows: Figure 3 As shown.
[0075] Step 2.1 includes: using max pooling to downsample the ground truth map of the original image size at different ratios to correspond to the output scale of each layer. The pooling kernel size and stride are set as the downsampling factor. While reducing the feature map size, the feature information of the crack key points is preserved. The feature value of each spatial location (i,j) in the feature map Y output after max pooling is calculated as shown in the following formula:
[0076] y ij =max(X (i×s:i×s+k,j×s:j×s+k) )
[0077] In the formula, X represents the input feature map, y ij represents the value of the output feature map at point (i,j), s represents the pooling kernel step size and feature map downsampling rate, and k represents the pooling kernel size.
[0078] The improved PolyHRNet model, used for crack keypoint detection, has only 9.64M parameters, while the U-net network, commonly used for crack semantic segmentation, has 28M parameters. In terms of the CKS evaluation metric, PolyHRNet achieves an accuracy of 0.81, a recall of 0.94, and a final F1 score of 0.87. This model can achieve 37fps on an RTX 2080s graphics card, while other lightweight segmentation models generally have inference speeds below 10fps on embedded devices. This model successfully achieves high accuracy and real-time performance requirements for crack detection while maintaining a lightweight design. Specifically, 9.64M represents the memory consumption of the model's parameters, obtained directly from the code output, indicating the model's lightweight nature; accuracy, recall, and F1 score are obtained using the CKS calculation formula, with the final F1 score indicating that high accuracy is met; 37fps represents 37 images processed per second, also obtained directly from the network code output, indicating that real-time performance is met.
[0079] In another embodiment, step 3 includes:
[0080] Step 3.1: The DBSCAN-based key point clustering algorithm is adopted to effectively identify irregularly shaped clusters and process noise data by analyzing the density distribution of the heat map. The sparse key points generated by the heat map clustering are connected, and the endpoints of the resulting multi-segment lines are matched after the connection is completed.
[0081] Step 3.2: Using the nearest neighbor-first iterative matching algorithm, a global distance matrix is established by iteratively calculating the Euclidean distance of each endpoint combination, and the two closest endpoints are selected in turn to connect them to form a continuous polyline segment;
[0082] Step 3.3: Exclude connected endpoints and iteratively update the distance matrix until the topology reconstruction of all key points is completed;
[0083] Step 3.4: Starting from any endpoint of the polyline, perform vector operation analysis on each pair of adjacent edges. The core detection parameter of the algorithm is defined as the angle θ between the direction vectors of adjacent edges and the angle α between the direction vectors of the interval. Determine whether the angle θ between the direction vectors of adjacent edges or the angle α between the direction vectors of the interval is greater than 90°. If it is greater, it is determined that there is an abnormal connection between the corresponding adjacent edges, and the abnormal connection is adjusted accordingly. Specifically, this includes:
[0084] Step 3.41: Detect the angle between the direction vectors of adjacent edges, θ. When θ > 90° is detected, mark the connection as an anomalous crack topology, such as... Figure 5 The angle θ formed by the BC and CD sides shown above prioritizes retaining the shorter side segments at abnormal connections, while deleting redundant long sides.
[0085] Step 3.42: For a composite connection structure consisting of three consecutive sides, namely sides AB, BC, and CD connected in sequence, calculate the included angle α between the extensions of the first and last sides AB and CD. If α > 90°, then the intermediate transition side BC is determined to be a redundant connection, and the side BC deletion operation is performed. Specifically, as follows... Figure 6 As shown, if α < 90°, then the intermediate transition edge BC is determined to be a non-redundant connection, specifically as follows: Figure 7 As shown.
[0086] In another embodiment, step 4 is included:
[0087] The dedicated evaluation metric used is CKS (Crack Keypoint Similarity) to evaluate the crack keypoint detection model. In CKS, the F1-score is calculated using precision and recall, and the formula for calculating the F1-score is as follows:
[0088]
[0089] In the formula, F 1_CKSRepresented as F1-score, P CKS Represented as spatial accuracy, R CKS Let P be the spatial recall rate. CKS and R CKS The specific calculation method is as follows:
[0090]
[0091] In the formula, N p G(p) represents the number of predicted keypoints. i ) represents the prediction key point p i In the actual heatmap, the activation value N g P(g) represents the number of predicted keypoints. i ) represents the prediction key point p i Activation values in a real heatmap;
[0092] Unlike the normal calculation methods for precision and recall, spatial precision is calculated by looking up the response intensity of the actual labeled points from the predicted keypoint coordinates. The actual value of the point on the actual heatmap corresponding to the predicted keypoint coordinates is used as the spatial precision. Spatial recall is calculated by looking up the response intensity of the predicted points from the predicted heatmap from the actual keypoint coordinates. The predicted value of the point on the predicted heatmap corresponding to the actual keypoint coordinates is used as the spatial recall.
[0093] In another embodiment, a bridge crack intelligent recognition system based on key points is disclosed, comprising: a dataset creation module, a model training module, a model evaluation module, and a recognition module. The dataset creation module is used to collect original crack images to create an original dataset, and then annotate the original dataset to obtain a crack dataset, the crack dataset including a crack ground truth map. The model training module is used to train an improved PolyHRNet network model using the crack dataset to obtain a crack key point detection model, inputting the original crack images into the crack key point detection model, and outputting a crack heatmap containing crack key points. The model evaluation module is used to evaluate the crack key point detection model using a dedicated evaluation metric CKS. The crack recognition module is used to obtain crack key points from the crack heatmap, and then connect the key points and generate polylines reflecting crack characteristics through post-processing operations.
[0094] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.
Claims
1. A method for intelligent identification of bridge cracks based on key points, characterized in that, Includes the following steps: Step 1: Collect original crack images to create an original dataset, and then label the original dataset to obtain a crack dataset, which includes crack ground truth maps; Step 2: Train the improved PolyHRNet network model using the crack dataset to obtain a crack key point detection model. Input the original crack image into the crack key point detection model and output a crack heat map containing crack key points. Step 3: After obtaining the key points of the crack from the crack heat map, the crack identification result is obtained by connecting the key points and generating polylines that reflect the crack characteristics through post-processing operations. The image processing of the crack key point detection model includes: Step 2.1: The original image is first passed through a Conv Block convolutional module with a downsampling factor of 4, and then through a layer1 convolutional module for feature extraction; Step 2.2: The feature map output in Step 2.1 continuously introduces new low-resolution branches in the process from stage 1 to stage 4, compresses the spatial dimension through downsampling operation, and adjusts the number of channels through 1x1 convolution; Step 2.3: In each stage, the high-resolution branch passes the features to the low-resolution branch through downsampling, and the low-resolution branch passes the semantic information back to the high-resolution branch through upsampling. Stage 3 is repeated four times and stage 4 is repeated three times, finally resulting in four output layers from head1 to head4. Step 2.4: After upsampling, head1 to head4 are stacked, then passed through the Neck module, and finally output through the head5 output layer to produce a heat map image containing the key points of the crack.
2. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, Step 1 includes: Step 1.1: Label the images in the original dataset. The original dataset is subjected to horizontal flipping, random scaling, vertical flipping, rotation and other operations to obtain the augmented crack dataset. The labeled multi-line segments are used to generate a crack ground truth map. The generated crack ground truth map is used as the new crack dataset. The crack ground truth map includes information on key points and crack connections, which is used for subsequent network model training. Step 1.2: Divide the crack dataset into training and validation sets for training the improved PolyHRNet network model.
3. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, The loss function used to train the improved PolyHRNet network model is the MSELoss function, and its calculation formula is shown in the following equation: ； In the formula, This represents the predicted value of the i-th sample at pixel position p. This represents the true value of the i-th sample at pixel position p. This represents the total number of key points. This indicates the number of categories of key points. Represents the spatial coordinates of the heatmap.
4. The intelligent bridge crack identification method based on key points according to claim 3, characterized in that, In the MSELoss loss function, the full loss calculation is retained for the zero-value region of the truth map, and the complete loss calculation is performed for the region of the truth map with a value greater than 0.
8. A loss threshold filtering mechanism of 0.04 is set. The specific calculation formula is as follows: 。 5. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, In step 2.4, the Neck module uses ordinary convolution, and then the predicted heatmap is obtained through 4 Basic Block residual modules.
6. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, Step 2.1 includes: using max pooling to downsample the ground truth map of the original image size at different ratios to correspond to the output scale of each layer. The pooling kernel size and stride are set as the downsampling factor. While reducing the feature map size, the feature information of the crack key points is preserved. The feature value of each spatial location (i, j) in the feature map Y output after max pooling is calculated as shown in the following formula: ； In the formula, X represents the input feature map, y ij represents the value of the output feature map at point (i, j), s represents the pooling kernel step size and feature map downsampling rate, and k represents the pooling kernel size.
7. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, Step 3 includes: Step 3.1: The DBSCAN-based key point clustering algorithm is adopted to effectively identify irregularly shaped clusters and process noise data by analyzing the density distribution of the heat map. The sparse key points generated by the heat map clustering are connected, and the endpoints of the resulting multi-segment lines are matched after the connection is completed. Step 3.2: Using the nearest neighbor-first iterative matching algorithm, a global distance matrix is established by iteratively calculating the Euclidean distance of each endpoint combination, and the two closest endpoints are selected in turn to connect them to form a continuous polyline segment; Step 3.3: Exclude connected endpoints and iteratively update the distance matrix until the topology reconstruction of all key points is completed; Step 3.4: Starting from any endpoint of the polyline, perform vector operation analysis on each pair of adjacent sides. The core detection parameter of the algorithm is defined as the angle θ between the direction vectors of adjacent sides and the angle α between the direction vectors of the interval. Determine whether the angle θ between the direction vectors of adjacent sides or the angle α between the direction vectors of the interval is greater than 90°. If it is greater, it is determined that there is an abnormal connection between the corresponding adjacent sides, and the abnormal connection is adjusted accordingly.
8. The intelligent bridge crack identification method based on key points according to claim 1, characterized in that, Including step 4: The dedicated evaluation metric used is CKS (Crack Keypoint Similarity) to evaluate the crack keypoint detection model. In CKS, the F1-score is calculated using precision and recall, and the formula for calculating the F1-score is as follows: ； In the formula, F 1_CKS Represented as F1-score, P CKS Represented as spatial accuracy, R CKS Let P be the spatial recall rate, where P CKS and R CKS The specific calculation method is as follows: ； ； In the formula, N p G(p) represents the number of predicted keypoints. i ) represents the predicted key point p. i In the actual heatmap, the activation value N g P(g) represents the number of predicted keypoints. i ) represents the predicted key point p. i Activation values in a real heatmap.
9. A bridge crack intelligent identification system based on key points, characterized in that, include: The system comprises a dataset creation module, a model training module, a model evaluation module, and a crack recognition module. The dataset creation module collects original crack images to create an original dataset, which is then labeled to obtain a crack dataset including ground truth crack images. The model training module trains an improved PolyHRNet network model using the crack dataset to obtain a crack keypoint detection model. The original crack images are input into the crack keypoint detection model, which outputs a crack heatmap containing crack keypoints. The model evaluation module evaluates the crack keypoint detection model using the CKS (Crack Detection and Evaluation) metric. The crack recognition module extracts crack keypoints from the crack heatmap and then uses post-processing to connect the keypoints and generate polylines reflecting crack characteristics. The image processing of the crack key point detection model includes: The original image is first passed through a Conv Block convolutional module with a downsampling factor of 4, and then through a layer 1 convolutional module for feature extraction; The output feature map continuously introduces new low-resolution branches in the process from stage 1 to stage 4, compresses the spatial dimension through downsampling operations, and adjusts the number of channels through 1x1 convolution. In each stage, the high-resolution branch passes the features to the low-resolution branch through downsampling, and the low-resolution branch passes the semantic information back to the high-resolution branch through upsampling. Stage 3 is repeated four times and stage 4 is repeated three times, finally resulting in four output layers from head1 to head4. After upsampling, head1 to head4 are stacked, then passed through the Neck module, and finally output through the head5 output layer to produce a heat map image containing the key points of the crack.