Intelligent recognition method and system for concrete cracks based on multi-segment line labeling
By combining polyline annotation and the U-Net network model, the problem of difficult dataset annotation in crack detection is solved, improving the efficiency and robustness of crack recognition. It is applicable to a variety of detection networks, especially for crack recognition in complex scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2025-05-15
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies for crack detection lack high-quality datasets and effective annotation methods, making crack identification difficult, especially in terms of robustness and accuracy in complex scenarios.
The existing concrete crack dataset is re-annotated using a polyline annotation method to generate a polyline dataset. The model is then trained using a U-Net network model. Through distance diffusion and loss function optimization, a ground truth map of the cracks is generated, enhancing the model's ability to learn high-dimensional semantic features.
It improves the annotation efficiency and robustness of crack identification, enhances the adaptability to complex scenarios, achieves higher identification accuracy and stability, and is applicable to a variety of detection networks.
Smart Images
Figure CN120543973B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of concrete structure detection technology, specifically relating to a method and system for intelligent identification of concrete cracks based on multi-segment line annotation. Background Technology
[0002] Due to the unique properties of concrete materials, identifying the causes of cracks, assessing their impact on bridge structures, and controlling cracks have always been research focuses. The causes and hazards of cracks vary across different structures. Large concrete structures like bridges have poor thermal conductivity, leading to heat accumulation in the core and a significant thermal gradient. This causes compressive stress in the core and tensile stress on the surface, resulting in surface cracks caused by the thermal gradient and external constraints. Currently, mainstream visual defect detection methods are mainly divided into two technical routes: object detection and semantic segmentation. Both methods are widely used in crack detection scenarios. Object detection technology, based on representative algorithms such as YOLO and Faster R-CNN, locates crack regions using bounding boxes. Its biggest advantage is its fast inference speed, making it suitable for online detection scenarios with high real-time requirements. Semantic segmentation technology employs a pixel-level classification strategy, classifying each pixel in an image as either a crack or background. The generated binary mask accurately represents the spatial distribution, topological structure, and detailed morphology of cracks.
[0003] However, there are no authoritative public datasets like COCO, ImageNet, and ADE20K in the field of crack detection, making it very difficult to obtain high-quality annotated crack semantic segmentation datasets. Most crack datasets have fewer than 600 images, and some even have fewer than 100. Many methods based on semi-supervised learning annotation and GAN generation have failed to fundamentally solve the problem of crack annotation difficulties. Summary of the Invention
[0004] To address the difficulty of crack annotation in existing technologies, this invention provides a method and system for intelligent identification of concrete cracks based on polysegment line annotation. The aim is to reduce annotation time while forcing the model to learn high-dimensional semantic features, enhancing the robustness of the algorithm in dealing with complex scenarios, providing a more reliable geometric feature basis for subsequent crack detection and quantitative analysis, and offering a new technical path for the health monitoring of concrete structures.
[0005] The technical solution adopted in this invention is as follows:
[0006] The intelligent identification method for concrete cracks based on polysegment line annotation includes the following steps:
[0007] S1: Using the existing publicly available concrete crack dataset as the base dataset, and using the polyline annotation method to re-annotate the crack features in the base dataset to obtain the polyline dataset;
[0008] S2: The U-Net network model is trained using the polyline dataset obtained in S1 to obtain a concrete crack intelligent recognition model based on polyline annotation.
[0009] S3: Using the concrete crack intelligent recognition model based on polyline annotation obtained from S2, concrete cracks are identified.
[0010] Preferably, the specific steps of S1 include:
[0011] S101: The X-AnyLabeling tool was used to label the original crack images in the basic dataset. The labeling format was a single-class polyline, the data format adopted the labelme polyline standard, and the data objects were stored in JSON format.
[0012] S102: Based on the annotation results obtained in S101, the ground truth map of the crack key points and the ground truth map of the crack connection line are generated by the distance diffusion method. The ground truth map of the crack key points and the ground truth map of the crack connection line are fused to obtain the fused ground truth map, which is the polyline dataset.
[0013] As a preferred option, the method of taking the maximum value is also used to fuse the true value map of the crack key points and the true value map of the crack connection obtained in S102, so as to use the true value of the loss function for calculation.
[0014] Specifically, the loss function used is the MSELoss function, which calculates the square of the difference between the predicted value and the true value. Its gradient is continuous, and it gradually decreases as the predicted value approaches the true value. The specific calculation formula is as follows:
[0015]
[0016] In the formula, N represents the sample size, y i Let the value be the true value of the i-th sample. Let L represent the predicted value of the i-th sample, and L be the loss value of this batch.
[0017] Preferably, the specific steps for generating the truth map of the crack key points in S102 include:
[0018] S1021: Perform data initialization. Create an all-zero floating-point matrix as the initial truth map based on the H×W size of the input image. Set the data type to float32 to ensure calculation accuracy. Traverse the key point array to perform geometric transformation. Convert the normalized relative coordinates to absolute pixel coordinates and set the label values at the corresponding positions in the truth map to form a binary key point distribution map.
[0019] S1022: Perform distance diffusion calculation. By calling the OpenCV cv2.distanceTransform function to perform Euclidean distance transformation, calculate the geometric distance of each pixel to the nearest key point, and expand the discrete point label into a continuous distribution with spatial decay characteristics.
[0020] S1023: Apply distance constraints and invert and normalize the distance matrix. The distance matrix is numerically clipped by setting the diffusion radius. Pixels outside the range are processed using boundary values. The processed matrix is normalized to the maximum and minimum values to ensure that the output value is within the set range. The original key point position maintains the response intensity of the maximum value of 1.
[0021] The specific steps for generating the truth map of crack connections are the same as those for generating the truth map of crack key points. When generating the truth map of crack connections, a distance diffusion coefficient is set, and an upper bound constraint of the response value threshold is adopted.
[0022] As a preferred approach, subjective judgment is introduced into the annotation process of S101 to enhance the network's ability to learn higher-order features. The annotation principle is to achieve the optimal fit of crack direction using the minimum number of key points.
[0023] Preferably, the model framework of the U-Net network model described in S2 includes:
[0024] The main body of the U-Net network includes encoder and decoder paths. Each encoder contains 4 repeating modules, and each module consists of two 3x3 convolutions + ReLU activations.
[0025] Downsampling is performed using 2x2 max pooling with a stride of 2, which halves the feature map size while doubling the number of channels. The encoder extracts high-level features through continuous convolution and pooling, gradually compressing the spatial dimension.
[0026] Upsampling, the decoder is symmetrical to the encoder and also contains 4 modules. Each module first doubles the feature map size and reduces the number of channels through transposed convolution. Then, the feature maps are stacked with the corresponding layer of the encoder through skip connections. The features are then refined by two 3x3 convolutions + ReLU.
[0027] Skip connections are introduced during the decoding and encoding process, where feature maps from each stage of the encoder are directly concatenated with the corresponding layers of the decoder.
[0028] As a preferred option, the process of training the U-Net network model using the polyline dataset obtained in S1 in S2 specifically includes:
[0029] The learning rate employs a dynamic adjustment strategy, and its calculation formula is as follows:
[0030]
[0031] In the formula, σ represents the learning rate, and η initial Let t represent the initial learning rate, droprate represent the decay rate, and t represent the current training step. max This is represented as the decay period;
[0032] The optimizer chooses stochastic gradient descent, which drives the variable term. The formula for calculating stochastic gradient descent is as follows:
[0033]
[0034] θ t =θ t-1 -η·v t
[0035] In the formula, v t Represented as the current velocity vector, v t-1 Let β represent the velocity vector at the previous moment, and let β represent the momentum coefficient. Represented as a loss function with respect to model parameters θ t-1 The gradient, θ t Represented as the updated model parameters;
[0036] The regularization method employs an L2 weight decay strategy, and its calculation formula is as follows:
[0037]
[0038] In the formula, J total J(θ) represents the total loss, J(θ) represents the original loss, λ represents the regularization coefficient, and θ represents the weight parameter. k This represents the k-th trainable parameter in the model.
[0039] As a preferred option, the evaluation of the crack intelligent recognition effect is also included, specifically through the following three metrics to assess and predict performance:
[0040] Precision: It is defined as the proportion of samples predicted as positive that are actually positive. The formula is:
[0041]
[0042] In the formula, TP represents the number of samples correctly predicted as positive by the model, and FP represents the number of negative samples incorrectly predicted as positive by the model.
[0043] Recall: It is defined as the proportion of samples that are actually positive that are correctly predicted by the model;
[0044]
[0045] In the formula, TP represents the number of samples correctly predicted as positive by the model, FP represents the number of negative samples incorrectly predicted as positive by the model, and FN represents the number of positive samples incorrectly predicted as negative by the model.
[0046] F1 score: It is defined as the harmonic mean of precision and recall;
[0047]
[0048] In the formula, p represents precision and r represents recall.
[0049] The intelligent concrete crack recognition system based on polysegment line annotation is characterized in that: the intelligent concrete crack recognition method based on polysegment line annotation includes the following modules:
[0050] The polyline annotation module is used to re-annotate the crack features in the existing publicly available concrete crack dataset using the polyline annotation method to obtain a polyline dataset.
[0051] The sample classification module is used to divide the multi-segment line dataset into training and test sets;
[0052] The model training module is used to train the U-Net network model using the training set to obtain a concrete crack intelligent recognition model based on polyline annotation.
[0053] The intelligent crack recognition module uses a concrete crack recognition model based on multi-segment line annotation to identify concrete cracks.
[0054] Preferably, it also includes an evaluation module for evaluating the model's performance.
[0055] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0056] 1. This invention proposes a crack identification method based on multi-segment line annotation. The annotation efficiency of this method is comparable to that of target detection methods, and the annotation time for a single sample can be controlled within 10-30 seconds, which has a significant advantage over semantic segmentation annotation methods (which usually require 10-30 minutes).
[0057] 2. Compared with traditional pixel-level annotation, the model based on polyline-generated ground truth map shows better generalization ability and stability in crack identification tasks. Its F1-Score is higher than that of the pixel-level annotation model under different thresholds. Polyline annotation can preserve the continuous morphological features of cracks, enabling the model to learn the geometric structure of cracks more fully and maintain a relatively stable recall rate under strict thresholds.
[0058] 3. The polyline annotation method described above can be trained not only on the U-Net network model, but also shows excellent performance in multiple mainstream detection networks such as YOLOv3, PPYOLO, and HRNet, depending on the user's task of cracking, including crack target localization or crack segmentation. Attached Figure Description
[0059] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:
[0060] Figure 1 This is a flowchart illustrating the crack identification method based on multi-segment line annotation of the present invention.
[0061] Figure 2 This is a schematic diagram of the U-Net network structure in this invention;
[0062] Figure 3 This is a schematic diagram of the original crack image labeled with polylines using X-AnyLabeling software in this invention;
[0063] Figure 4 This is a schematic diagram of generating a crack truth map from an annotated image in this invention;
[0064] Figure 5 This describes the effect of different segmentation algorithms on crack segmentation in this invention. Detailed Implementation
[0065] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0066] The intelligent identification method for concrete cracks based on polysegment line annotation includes the following steps:
[0067] S1: The existing publicly available concrete crack dataset is used as the base dataset (from Volume 234 of Building and Building Materials, "Image-Based Concrete Crack Detection in Tunnels"). This dataset contains 919 images and corresponding pixel-level annotation files. The crack features in the base dataset are re-annotated using the polyline annotation method to obtain the polyline dataset. The specific process is as follows:
[0068] S101: First, the basic dataset is classified using a binary classification method to obtain a training set and a test set, with 735 samples in the training set and 184 samples in the test set. The image resolution in the basic dataset is 512×512 pixels. A bilinear downsampling method is used to uniformly adjust the test set images and their corresponding pixel-level labeled images to 256×256 pixels. Then, the X-AnyLabeling tool is used to label the original crack images in the basic dataset (e.g., ...). Figure 3 As shown), the annotation form is a single-class polyline, the data format adopts the labelme polyline standard, and the data object is stored in JSON format; in the annotation process, in addition to focusing on the morphological features of the crack, subjective judgment is also introduced to enhance the network's ability to learn higher-order features and improve the robustness of the model. The annotation principle is to use the minimum number of key points to achieve the optimal fit of the crack direction.
[0069] S102: Based on the annotation results obtained in S101, ground truth maps of crack key points and crack lines are generated using the distance diffusion method. These ground truth maps are then fused to obtain a fused ground truth map, which is the polyline dataset (containing the ground truth map generated based on the polyline and the original image). The specific process includes:
[0070] S1021: Perform data initialization. Create an all-zero floating-point matrix as the initial truth map based on the H×W size of the input image. Set the data type to float32 to ensure calculation accuracy. Traverse the key point array to perform geometric transformation. Convert the normalized relative coordinates to absolute pixel coordinates and set the label values at the corresponding positions in the truth map to form a binary key point distribution map.
[0071] S1022: Perform distance diffusion calculation. By calling the OpenCV cv2.distanceTransform function to perform Euclidean distance transformation, calculate the geometric distance of each pixel to the nearest key point, and expand the discrete point label into a continuous distribution with spatial decay characteristics.
[0072] S1023: Apply distance constraints and invert and normalize the distance matrix. The distance matrix is numerically clipped by setting the diffusion radius. Pixels outside the range are processed using boundary values. The processed matrix is normalized to the maximum and minimum values to ensure that the output value is within the set range. The original key point position maintains the response intensity of the maximum value of 1.
[0073] The same operations from S1021 to S1023 are used to generate a truth map of the crack connection (e.g.) Figure 4 As shown in the figure, the difference lies in setting the distance diffusion coefficient and adopting a strict upper bound constraint on the response value threshold (the maximum value is limited to 0.8), which effectively highlights the importance of the marked key points and reasonably maintains the characteristics of crack path continuity.
[0074] It also includes using the maximum value method to fuse the true value map of the crack key points and the true value map of the crack connection obtained in S102, so as to use it as the true value of the loss function;
[0075] Specifically, the loss function used is the MSELoss function, which calculates the square of the difference between the predicted value and the true value. Its gradient is continuous, and it gradually decreases as the predicted value approaches the true value. The specific calculation formula is as follows:
[0076]
[0077] In the formula, N represents the sample size, y i Let the value be the true value of the i-th sample. Let L represent the predicted value of the i-th sample, and L be the loss value of this batch.
[0078] S2: The U-Net network model is trained using the polyline dataset obtained in S1 to obtain a concrete crack intelligent recognition model based on polyline annotation.
[0079] like Figure 2 As shown, the model framework of the U-Net network model includes:
[0080] The main body of the U-Net network includes encoder and decoder paths. Each encoder contains 4 repeating modules, and each module consists of two 3x3 convolutions + ReLU activations.
[0081] Downsampling is performed using 2x2 max pooling with a stride of 2, which halves the feature map size while doubling the number of channels. The encoder extracts high-level features through continuous convolution and pooling, gradually compressing the spatial dimension.
[0082] Upsampling, the decoder is symmetrical to the encoder and also contains 4 modules. Each module first doubles the feature map size and reduces the number of channels through transposed convolution. Then, the feature maps are stacked with the corresponding layer of the encoder through skip connections. The features are then refined by two 3x3 convolutions + ReLU.
[0083] Skip connections are introduced during the decoding and encoding process, where feature maps from each stage of the encoder are directly concatenated with the corresponding layers of the decoder.
[0084] The model runs for 160 epochs. The batch size is dynamically adjusted to 4 based on GPU memory usage. The learning rate uses a dynamic adjustment strategy, with an initial learning rate set at 0.0001. When the number of training iterations reaches 10,000 and 20,000, the learning rate is decayed in a stepwise manner with a coefficient of 0.1, as shown in the following formula:
[0085]
[0086] In the formula, η represents the learning rate. initial Let t represent the initial learning rate, droprate represent the decay rate, and t represent the current training step. max This is represented as the decay period;
[0087] The optimizer uses stochastic gradient descent (SGF) with a momentum coefficient of 0.9. By preserving components of historical gradient directions during parameter updates, it effectively accelerates the convergence of the model in the optimization direction while suppressing gradient oscillations. The formula for calculating stochastic gradient descent is as follows:
[0088]
[0089] θ t =θ t-1 -η·v t
[0090] In the formula, v t Represented as the current velocity vector, v t-1 Let β represent the velocity vector at the previous moment, and let β represent the momentum coefficient. Represented as a loss function with respect to model parameters θ t-1 The gradient, θ t Represented as the updated model parameters;
[0091] The regularization method employs an L2 weight decay strategy with a regularization coefficient set to 0.0005. By introducing a squared penalty term for the weight parameters into the loss function, it progressively suppresses large weight values in the network, thereby reducing the risk of model overfitting. The calculation formula is as follows:
[0092]
[0093] In the formula, J total J(θ) represents the total loss, J(θ) represents the original loss, λ represents the regularization coefficient, and θ represents the weight parameter. k This represents the k-th trainable parameter in the model;
[0094] Furthermore, the loss function used is the MSELoss function, which calculates the square of the difference between the predicted value and the true value. Its gradient is continuous, and it gradually decreases as the predicted value approaches the true value. The specific calculation formula is as follows:
[0095]
[0096] In the formula, N represents the sample size, y i Let the value be the true value of the i-th sample. This is represented as the predicted value of the i-th sample;
[0097] S3: Using the concrete crack intelligent recognition model based on polyline annotation obtained from S2, concrete cracks are identified.
[0098] In one embodiment, the evaluation of the crack intelligent recognition effect is also included. In this invention, the existing publicly available concrete crack dataset mentioned in S1 (which is also divided into a training set of 735 samples and a test set of 184 samples) is used to directly train and test the U-Net network model (using the test set) to obtain a control model. Then, the predictive performance of the two models is evaluated through the following three quality metrics:
[0099] Precision: It is defined as the proportion of samples predicted as positive that are actually positive. The formula is:
[0100]
[0101] In the formula, TP represents the number of samples correctly predicted as positive by the model, and FP represents the number of negative samples incorrectly predicted as positive by the model.
[0102] Recall: It is defined as the proportion of samples that are actually positive that are correctly predicted by the model;
[0103]
[0104] In the formula, TP represents the number of samples correctly predicted as positive by the model, FP represents the number of negative samples incorrectly predicted as positive by the model, and FN represents the number of positive samples incorrectly predicted as negative by the model.
[0105] F1 score: It is defined as the harmonic mean of precision and recall;
[0106]
[0107] In the formula, p represents precision and r represents recall.
[0108] Because each model employs two different ground truth map generation paradigms during training—pixel-level and polyline annotation—their evaluation metrics differ fundamentally, and direct comparison would lead to distorted conclusions. To establish a comparable model evaluation system, this invention unifies the final evaluation of both types of models to a pixel-level annotation benchmark. Pixel-level annotation, as an industry standard, has annotation accuracy approaching the limits of human judgment, and this annotation method can objectively reflect the model's fine-grained segmentation capability in complex backgrounds.
[0109] The experiment employed a dynamic threshold scanning method, conducting systematic testing within a threshold range of 0.1-0.9 with a step size of 0.1. The results of two network models trained on different training sets were compared, as shown in Table 1.
[0110] Table 1 Evaluation of segmentation results of different training set models
[0111]
[0112] It is evident that the model based on polyline-generated ground truth maps exhibits superior generalization ability and stability in the crack identification task. Its F1-Score fluctuates between 0.27 and 0.36 and is generally higher than the pixel-annotated model, especially maintaining an F1 value of approximately 0.30-0.34 at higher thresholds (0.5-0.9), while the pixel-annotated model's F1 value drops sharply to below 0.20 after the threshold exceeds 0.4. This suggests that polyline annotation may preserve the continuous morphological features of cracks, allowing the model to learn the crack's geometric structure more thoroughly and maintain a relatively stable recall rate even under strict thresholds. While the pixel-annotated model demonstrates higher precision at medium to low thresholds (0.1-0.3), reaching a maximum of 0.41, its drastically decreased recall rate reveals a serious problem of missed detections, possibly stemming from overfitting due to the excessive sensitivity of pixel-level annotations to subtle texture changes.
[0113] The effects of different segmentation algorithms on crack segmentation, such as Figure 5 As shown, by visually comparing the ground truth annotations and the model output heatmap, it can be found that the model trained based on pixel annotations exhibits significant feature response decay in areas with blurred network features, and the activation values in the heatmap are generally below the discrimination threshold of 0.3. This phenomenon may be because the proportion of obvious crack samples in the dataset is too high, while the number of crack samples with blurred features is too low. The crack's morphology, direction, lighting conditions, and background complexity lack diversity, making it difficult for the model to generalize to less obvious scenes. In contrast, the model trained based on polyline annotations can effectively capture the structural features of cracks.
[0114] Further analysis was conducted using thresholds that maximized the F1 scores of both the polyline-based and pixel-based models. The threshold for the polyline-based model was 0.2, while the threshold for the pixel-based model was 0.1. In 184 test samples, the average F1 score of the pixel-annotated model was 0.33, slightly lower than the 0.36 of the polyline-based model. The pixel-annotated model also exhibited greater performance volatility, with more frequent extreme high and low values. The accuracy of pixel annotation resulted in an F1 score as high as 0.80, higher than the 0.76 of the polyline-based model, particularly evident in samples with complex crack morphologies and rich details. This advantage stemmed from the detailed characterization of edges and minute cracks by the pixel-level ground truth map. However, the median F1 score of the pixel-annotated model was only 0.36, lower than the 0.38 of the polyline-based model. Polyline annotation showed greater stability in some samples, especially when the annotations contained noise or the crack continuity was poor; its geometrically based ambiguity may have reduced the risk of overfitting. It is worth noting that the F1 score of the pixel-annotated model includes 14 zero values, reflecting its sensitivity to annotation noise and its tendency to fail on low-contrast or fractured surfaces. The polyline annotation model, on the other hand, exhibits stronger robustness, with only 6 zero-value samples, and maintains comparable performance in scenes with strong crack continuity. This is related to the regularized ground truth map generated through polygon smoothing. However, due to the simplification of the annotation geometry, the F1 score of samples with complex topologies lags behind that of the pixel-annotated model. The two models perform similarly in 20% of the samples, mainly concentrated on moderately difficult cracks.
[0115] 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 concrete cracks based on polysegment line annotation, characterized in that: Includes the following steps: S1: Using an existing publicly available concrete crack dataset as the base dataset, and employing polyline annotation to re-annotate the crack features in the base dataset, a polyline dataset is obtained. Specific steps include: S101: The X-AnyLabeling tool was used to label the original crack images in the basic dataset. The labeling format was a single-class polyline, the data format adopted the labelme polyline standard, and the data objects were stored in JSON format. S102: Based on the annotation results obtained in S101, the ground truth map of the crack key points and the ground truth map of the crack connection line are generated by the distance diffusion method. The ground truth map of the crack key points and the ground truth map of the crack connection line are fused to obtain the fused ground truth map, which is the polyline dataset. Subjective judgment is introduced in the S101 annotation process to enhance the network's ability to learn higher-order features. The annotation principle is to achieve the optimal fit of crack direction using the minimum number of key points. It also includes using the maximum value method to fuse the true value map of the crack key points and the true value map of the crack connection obtained in S102, so as to use it as the true value of the loss function; Specifically, the loss function used is the MSELoss function, which calculates the square of the difference between the predicted value and the true value. Its gradient is continuous, and it gradually decreases as the predicted value approaches the true value. The specific calculation formula is as follows: ； In the formula, N This represents the number of samples in a single batch. Represented as the first The true value of each sample Represented as the first The predicted value for each sample, where L is the loss value for that batch; The specific steps for generating the truth map of crack key points in S102 include: S1021: Perform data initialization. Create an all-zero floating-point matrix as the initial truth map based on the H×W size of the input image. Set the data type to float32 to ensure calculation accuracy. Traverse the key point array to perform geometric transformation. Convert the normalized relative coordinates to absolute pixel coordinates and set the label values at the corresponding positions in the truth map to form a binary key point distribution map. S1022: Perform distance diffusion calculation. By calling the OpenCV cv2.distanceTransform function to perform Euclidean distance transformation, calculate the geometric distance of each pixel to the nearest key point, and expand the discrete point label into a continuous distribution with spatial decay characteristics. S1023: Apply distance constraints and invert and normalize the distance matrix. The distance matrix is numerically clipped by setting the diffusion radius. Pixels outside the range are processed using boundary values. The processed matrix is normalized to the maximum and minimum values to ensure that the output value is within the set range. The original key point position maintains the response intensity of the maximum value of 1. The specific steps for generating the truth map of crack connections are the same as those for generating the truth map of crack key points. When generating the truth map of crack connections, a distance diffusion coefficient is set, and an upper bound constraint of the response value threshold is adopted. S2: The U-Net network model is trained using the polyline dataset obtained in S1 to obtain a concrete crack intelligent recognition model based on polyline annotation. S3: Using the concrete crack intelligent recognition model based on polyline annotation obtained from S2, concrete cracks are identified.
2. The intelligent concrete crack identification method based on polysegment line annotation according to claim 1, characterized in that: In S1, a binary classification method is used to classify the polyline dataset to obtain the training set and the test set. The bilinear downsampling method is used to uniformly adjust the test set images and their corresponding pixel-level labeled images to 256×256 pixels. Then, the images in the test set are re-labeled using the polyline labeling method.
3. The intelligent concrete crack identification method based on polysegment line annotation according to claim 1, characterized in that: The model framework of the U-Net network model described in S2 includes: The main body of the U-Net network includes encoder and decoder paths. Each encoder contains 4 repeating modules, and each module consists of two 3x3 convolutions + ReLU activations. Downsampling is performed using 2x2 max pooling with a stride of 2, which halves the feature map size while doubling the number of channels. The encoder extracts high-level features through continuous convolution and pooling, gradually compressing the spatial dimension. Upsampling, the decoder is symmetrical to the encoder and also contains 4 modules. Each module first doubles the feature map size and reduces the number of channels through transposed convolution. Then, the feature maps are stacked with the corresponding layer of the encoder through skip connections. The features are then refined by two 3x3 convolutions + ReLU. Skip connections are introduced during the decoding and encoding process, where feature maps from each stage of the encoder are directly concatenated with the corresponding layers of the decoder.
4. The intelligent concrete crack identification method based on polysegment line annotation according to claim 1, characterized in that: The process of training the U-Net network model using the polyline dataset obtained in S1 specifically includes: The learning rate employs a dynamic adjustment strategy, and its calculation formula is as follows: ； In the formula, Represented as the learning rate, Let t represent the initial learning rate, droprate represent the decay rate, and t represent the current training step. max This is represented as the decay period; The optimizer chooses stochastic gradient descent, which drives the variable term. The formula for calculating stochastic gradient descent is as follows: ； ； In the formula, Represented as the current velocity vector, Represented as the velocity vector at the previous moment, Expressed as momentum coefficient, Represented as a loss function on model parameters gradient, Represented as the updated model parameters; The regularization method employs an L2 weight decay strategy, and its calculation formula is as follows: ； In the formula, Expressed as total loss, Represented as the original loss, Represented as the regularization coefficient, Represented as weight parameters, θ k The first in the model k There are 1 trainable parameters.
5. The intelligent concrete crack identification method based on polysegment line annotation according to claim 1, characterized in that: It also includes an evaluation of the effectiveness of intelligent crack recognition, specifically using the following three metrics to assess predictive performance: Precision: It is defined as the proportion of samples predicted as positive that are actually positive. The formula is: ； In the formula, TP This represents the number of samples that the model correctly predicted as positive. FP This represents the number of negative class samples that the model incorrectly predicted as positive. Recall: It is defined as the proportion of samples that are actually positive that are correctly predicted by the model; ； In the formula, TP This represents the number of samples that the model correctly predicted as positive. FP This represents the number of negative class samples that the model incorrectly predicted as positive. FN This represents the number of positive class samples that the model incorrectly predicted as negative. F1 Score: It is defined as the harmonic mean of precision and recall; ； In the formula, p Expressed as accuracy, r This is expressed as recall rate.
6. A concrete crack intelligent recognition system based on polysegment line annotation, characterized in that: The method for intelligent identification of concrete cracks based on polysegment line annotation as described in any one of claims 1-5 includes the following modules: The polyline annotation module is used to re-annotate the crack features in the existing publicly available concrete crack dataset using the polyline annotation method to obtain a polyline dataset. The sample classification module is used to divide the multi-segment line dataset into training and test sets; The model training module is used to train the U-Net network model using the training set to obtain a concrete crack intelligent recognition model based on polyline annotation. The intelligent crack recognition module uses a concrete crack recognition model based on multi-segment line annotation to identify concrete cracks.