A lightweight concrete structure crack automatic detection method

Abstract

The present application provides a kind of lightweight concrete structure crack automatic detection method.The lightweight concrete structure crack automatic detection method, S1, at least one high-resolution camera is used to capture the image of concrete structure under multiple illumination conditions, S2, the captured image is preprocessed to enhance the crack visibility under low light conditions, S3, the backbone network based on crack automatic detection network V3 and bottleneck transformer is used for feature extraction of image, S4, feature pyramid network and path aggregation network are applied to multi-scale fusion of features.The lightweight concrete structure crack automatic detection method, the lightweight concrete structure crack automatic detection method, a lightweight crack feature automatic extraction network is built, the network uses crack automatic detection network V3 and transformer, the addition of transformer enhances the crack feature extraction effect, and the introduction of channel attention mechanism makes the network pay more attention to crack features.

Description

Technical Field

[0001] This invention relates to the field of automatic crack detection technology for concrete structures, specifically to an automatic crack detection method for lightweight concrete structures. Background Technology

[0002] In transportation infrastructure, cracking is one of the common failure modes. After prolonged or excessive use, most pavement defects first manifest as cracks. If not repaired promptly, these cracks will continue to widen and extend due to external environmental or human factors. Driving at high speeds in such conditions not only affects driving comfort but may even lead to serious traffic accidents. The continuous expansion of even small cracks can pose potential dangers. Repairing cracks before they worsen can significantly reduce maintenance budgets, prevent accidents, and extend the road's lifespan. Crack detection has become a crucial part of concrete structure maintenance in order to assess the safety and durability of concrete structures. Traditional crack detection relies on manual visual inspection or uses methods such as electromagnetic, infrared, ultrasonic, or radioactive detection. Manual visual inspection is not only inefficient and costly but also has geographical limitations; for example, it is unsuitable for areas such as elevated bridge surfaces and piers. While external equipment-based detection methods have improved and can achieve certain expected results, these methods still involve some subjective factors.

[0003] Current image-based deep learning crack detection methods require large amounts of data to train the algorithm model. Collecting crack data and uploading it to the computing end is a time-consuming process, and cloud computing has high latency, making it impossible to respond quickly to incidents caused by cracks. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a lightweight automatic crack detection method for concrete structures. This method solves the problems of time-consuming processes such as collecting crack data and uploading it to the computing end, the high latency of cloud computing which makes it impossible to respond quickly to crack-related accidents, and the inability to extract crack features even under poor lighting conditions.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: 1. An automatic detection method for cracks in lightweight concrete structures, characterized in that it includes:

[0008] S1. Use at least one high-resolution camera to capture images of the concrete structure under various lighting conditions;

[0009] S2. Preprocess the captured images to enhance crack visibility under low light conditions;

[0010] S3. Feature extraction of the image is performed using the V3 crack detection network and the backbone network of Bottleneck Transformer. By adopting the V3 network and the Bottleneck Transformer structure, this scheme achieves lightweight and efficient feature extraction, which allows crack features to be better preserved in the deep network structure. The self-attention mechanism of Bottleneck Transformer improves the accuracy of feature extraction and reduces the interference of redundant features.

[0011] S4. Multi-scale feature fusion is performed using a feature pyramid network and a path aggregation network. A parameter-free attention mechanism, SimAM, is inserted into the Neck network. Ablation experiments demonstrate that this module improves feature fusion. Finally, the fused feature map is fed into the prediction end to generate many candidate anchor boxes. Non-maximum suppression is used to eliminate redundant prediction boxes. By setting a threshold, non-maximum suppression selects the prediction box with the highest score and deletes those prediction boxes that significantly overlap with it. The score is measured by IoU. During the candidate box generation process, non-maximum suppression combined with joint intersection algorithm is used for accurate localization. This combination effectively eliminates redundant prediction boxes while controlling the accuracy of crack recognition through IoU threshold, making the final detection results more reliable.

[0012] S5. Use nonmaximum suppression and joint intersection indices for precise crack location and identification;

[0013] S6. Deploy the model on an edge computing device to enable real-time crack detection.

[0014] Preferably, the image preprocessing step includes automatic exposure adjustment and contrast enhancement, and the feature extraction step includes applying a multi-head attention mechanism to emphasize crack features. The model deployed on the edge computing device has real-time data processing capabilities, and the response latency for crack detection is less than 50ms, ensuring rapid identification and response in harsh environments, thereby achieving more efficient crack detection under field conditions.

[0015] Preferably, the feature fusion step includes using a parameterless attention mechanism to enhance the representation of key crack features.

[0016] Preferably, the crack localization and identification step includes applying continuous intersection hyperjoint as a localization loss function.

[0017] Preferably, the attention mechanism to enhance the expression of key crack features is defined as:

[0018]

[0019] Where S is the attention map, x ij is a pixel in the feature map, and f represents the transformation function applied to each pixel.

[0020] Preferably, the feature fusion strategy utilizes a combination of a feature pyramid network and a path aggregation network. The feature pyramid network is responsible for extracting features at different depth levels, while the path aggregation network is responsible for more efficiently fusing and transmitting these features, ensuring that the flow of information from lower to higher levels is unimpeded, thereby optimizing the entire network's ability to detect cracks.

[0021] Preferably, the model deployed on the edge computing device has real-time data processing capabilities, rapid response, and instant updates.

[0022] (III) Beneficial Effects

[0023] This invention provides an automatic crack detection method for lightweight concrete structures. It has the following beneficial effects:

[0024] This lightweight automatic crack detection method for concrete structures constructs a lightweight crack feature extraction network. This network utilizes the Crack Detection Network V3 and the Transformer. The addition of the Transformer significantly enhances the crack feature extraction effect, and the introduction of a channel attention mechanism makes the network more focused on crack features. The path aggregation network is improved using the Transformer, further enhancing the feature fusion effect. Tests were conducted on images containing road surface cracks, and the results show that the inference time for a single image is only 11.3 ms. The lightweight automatic crack detection network not only reduces the computer load but also improves the speed and accuracy of crack detection, verifying the robust performance of the algorithm model. The automatic crack detection model can detect road surface cracks not only under normal conditions but also under poor lighting conditions and rainy weather conditions. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the fusion process of the feature pyramid and path aggregation network of the present invention;

[0026] Figure 2 This refers to the SimAM attention mechanism. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figures 1-2 As shown, this embodiment of the invention provides a lightweight automatic crack detection method for concrete structures, including, S1, using at least one high-resolution camera to capture images of the concrete structure under various lighting conditions.

[0029] S2. Preprocess the captured image to enhance crack visibility under low light conditions. The image preprocessing steps include automatic exposure adjustment and contrast enhancement, and the feature extraction steps include applying a multi-head attention mechanism to emphasize crack features.

[0030] S3. Image feature extraction is performed using the V3 crack detection network and the backbone network of the Bottleneck Transformer. The feature fusion strategy combines a feature pyramid network and a path aggregation network. The feature pyramid network is responsible for extracting features at different depth levels, while the path aggregation network is responsible for more efficient fusion and transmission of these features, ensuring unimpeded information flow from lower to higher layers, thereby optimizing the network's crack detection capability. In the feature pyramid, since it is a process of moving from large feature maps to small feature maps and then from small feature maps to large feature maps, shallow feature information is severely lost after multiple layers of transmission. Therefore, a bottom-up path aggregation network is added after the feature pyramid. The i-th feature map of the path aggregation network is fused with the (i+1)-th feature map of the feature pyramid to obtain a new fused feature map. This method enhances the representational ability of the features. By adopting the V3 network and the Bottleneck Transformer structure, this scheme achieves lightweight and efficient feature extraction, allowing crack features to be better preserved in deep network structures. The self-attention mechanism of the Bottleneck Transformer improves the accuracy of feature extraction and reduces the interference of redundant features.

[0031] S4. Multi-scale feature fusion is performed using a feature pyramid network and a path aggregation network. A parameter-free attention mechanism, SimAM, is inserted into the Neck network. Ablation experiments demonstrate that this module improves feature fusion. The fused feature map is then fed into the prediction end, generating many candidate anchor boxes. Non-maximum suppression (NMS) is used to eliminate redundant prediction boxes. By setting a threshold, NMS selects the prediction box with the highest score, while deleting those that significantly overlap with it. The score is measured by IoU. The feature fusion step includes using a parameter-free attention mechanism to enhance the expression of key crack features. This attention mechanism, defined as:

[0032]

[0033] Where S is the attention map, x ij is the pixel in the feature map, and f represents the transformation function applied to each pixel. During the candidate box generation process, nonmaximum suppression combined with joint intersection algorithm is used for accurate localization. This combination effectively eliminates redundant prediction boxes while controlling the accuracy of crack identification through IoU threshold, making the final detection result more reliable.

[0034] S5. Use nonmaximum suppression and joint intersection index for precise crack localization and identification. The crack localization and identification steps include applying continuous intersection superjoint as the localization loss function.

[0035] S6. Deploy the model on an edge computing device to achieve real-time crack detection. The model deployed on the edge computing device has real-time data processing capabilities, rapid response and instant update capabilities. The model deployed on the edge computing device has real-time data processing capabilities, and the response latency for crack detection is less than 50ms, ensuring rapid identification and response in harsh environments, thereby achieving more efficient crack detection under field conditions.

[0036] Model training and validation:

[0037] The model training epochs were set to 600, the batch size to 8, the initial learning rate to 0.01, and the weight decay to 0.0005. We used a stochastic gradient descent optimization strategy. Due to the small initial learning rate, we also used a cosine annealing learning rate decay strategy until the network parameters stabilized during training before accelerating the training. As training progressed, the model gradually approached the optimal solution. At this point, we reduced the learning rate updates, making the training results infinitely close to the optimal solution. With the continuous increase in training iterations, both the localization loss and confidence loss decreased and gradually stabilized until the loss values ​​almost stopped changing, indicating model convergence. The final training localization loss and confidence loss approached 0.045 and 0.072, respectively. In machine learning and deep learning, an epoch refers to the process of using the entire training dataset completely once during training. When the model is trained, it performs forward and backward propagation on the training set to adjust the model weights; this is called a training step or iteration.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automatic crack detection method for lightweight concrete structures, characterized in that, include: S1. Use at least one high-resolution camera to capture images of the concrete structure under various lighting conditions; S2. Preprocess the captured images to enhance crack visibility under low light conditions; S3. Feature extraction of the image is performed using the V3 crack detection network and the backbone network of Bottleneck Transformer. By adopting the V3 network and Bottleneck Transformer structure, crack features can be preserved in the deep network structure. The self-attention mechanism of Bottleneck Transformer improves the accuracy of feature extraction and reduces the interference of redundant features. S4. Multi-scale feature fusion is performed using a feature pyramid network and a path aggregation network. A parameter-free attention mechanism, SimAM, is inserted into the Neck network. Finally, the fused feature map is fed into the prediction end to generate many candidate anchor boxes. Non-maximum suppression is used to eliminate redundant prediction boxes. By setting a threshold, non-maximum suppression selects the prediction box with the highest score and deletes those prediction boxes that significantly overlap with this prediction box. The score is measured by IoU. During the candidate box generation process, non-maximum suppression combined with the joint intersection algorithm is used for accurate localization. The above method effectively eliminates redundant prediction boxes while controlling the accuracy of crack recognition through IoU threshold. S5. Use nonmaximum suppression and joint intersection indices for precise crack location and identification; S6. Deploy the methods of S1-S5 to achieve real-time crack detection. After deployment on the edge computing device, the above methods have real-time data processing capabilities, and the response latency of crack detection is less than 50ms. The feature fusion step includes using a parameterless attention mechanism to enhance the representation of key crack features. The crack localization and identification step includes applying a continuous intersection hyperjoint as a localization loss function. The attention mechanism to enhance the representation of key crack features is defined as follows: Where S is the attention map, x ij is a pixel in the feature map, and f represents the transformation function applied to each pixel. The feature fusion strategy utilizes a combination of a feature pyramid network and a path aggregation network. The feature pyramid network is responsible for extracting features at different depth levels, while the path aggregation network is responsible for fusing and transmitting these features to ensure that the information flow from lower to higher levels is unimpeded, thereby optimizing the entire network's ability to detect cracks.

2. The automatic crack detection method for lightweight concrete structures according to claim 1, characterized in that: The image preprocessing steps include automatic exposure adjustment and contrast enhancement, and the feature extraction steps include applying a multi-head attention mechanism to emphasize crack features.

3. The automatic crack detection method for lightweight concrete structures according to claim 1, characterized in that: The model deployed on the edge computing device has real-time data processing capabilities, rapid response, and instant updates.

Images