A dam crack detection method based on intelligent unmanned aerial vehicle

Abstract

The application provides a dam crack detection method based on an intelligent unmanned aerial vehicle, comprising the following steps: 1. Data set construction: using an unmanned aerial vehicle to shoot crack pictures or record videos of the dam surface under real engineering conditions, frame the shot videos to obtain crack images, and obtain a data set; 2. Dam crack detection model construction; 3. Dam crack detection model weight conversion; 4. Edge deployment of a computing processor: realizing image burning of the computing processor on a host computer, and then migrating and deploying the model processed in step 3; 5. Dam surface crack detection by an unmanned aerial vehicle: carrying the computing processor on the intelligent unmanned aerial vehicle, directly inputting the shot video images into the deployed computing processor for processing during aerial photography of the intelligent unmanned aerial vehicle, and detecting the dam surface cracks by GE-YOLOv5. The detection method improves the accuracy and efficiency of crack detection by modifying YOLOv5, and can realize real-time detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of engineering detection technology, and particularly relates to a dam crack detection method based on an intelligent unmanned aerial vehicle. BACKGROUND

[0002] A dam is an important part of water conservancy construction in China, and has the functions of flood control, power generation, irrigation, water supply and navigation. Once the dam is broken or collapses, it will cause serious harm to the downstream and surrounding areas and cause huge losses to the national economy. The key factor affecting the safety of the dam is the crack. The crack is a special image feature that appears in the form of a thin and irregular black line shadow, and is usually hidden in the surface of the dam. This is also the difficulty of dam surface crack detection. In the crack detection process, it is easy to be affected by various noise backgrounds, resulting in inaccurate detection results. Therefore, it is necessary to improve the current target detection algorithm and conduct in-depth research on dam surface crack detection to make up for the shortcomings of the existing algorithm.

[0003] Through the search of the existing technology, literature and patents, it is found that most of the existing dam surface crack detection methods are through traditional manual visual inspection, manual interpretation or through a deep learning target detection algorithm without improvement to find cracks. However, the manual identification method is low in efficiency, and the use of manpower alone cannot guarantee all-around coverage in terms of time, technology and funds. Moreover, it is greatly affected by subjective factors, and it is likely to miss and eliminate safety hazards. With the development of computer vision, image processing and machine vision technology have been applied in many fields. Traditional image processing techniques, such as crack edge detection algorithms, mainly include Sobel edge detection operator, Canny edge detection operator and Laplacian operator. However, after consulting a large number of literature, these algorithms are only suitable for the detection of cracks with large width, simple background and obvious cracks. Therefore, the traditional crack edge detection algorithm has low robustness and is not suitable for images with hidden cracks and variable backgrounds. When the crack image background is complex and noisy, the detection accuracy of the traditional edge detection method decreases significantly. Most of the deep learning convolutional neural networks are used to learn and classify crack images, and a target detection model is used to detect and identify cracks. According to the investigation, the general target detection algorithm based on deep learning is more applied to the detection of road surfaces and bridges. However, the dam crack appears in the form of a thin and irregular black line shadow, which is difficult to detect. Moreover, the crack concrete and the surrounding environment are very similar, resulting in low detection accuracy of the general target detection algorithm for dam surface crack detection, and many missed detections.

[0004] YOLOv5 is a high-efficiency and fast single-stage target detection algorithm, although the YOLOv5 target detection model has been applied in many fields, but if it is directly applied to the surface crack detection of the dam surface, the following problems still exist: (1) the dam crack is in the form of a thin and irregular black line shadow, and the common model is difficult to identify and has low precision. (2) The dam crack and the surrounding environment are very similar, the accuracy is low, and the missed detection is easy to produce; (3) at present, the unmanned aerial vehicle shoots the video image of the dam surface, and then the shot video image is imported into the deep learning workstation for processing, and the detection processing efficiency is low. (4) The model weight obtained after general model training is large, and cannot be well embedded in the edge AI computing processor Jetson XavierTM NX. (5) The computing capacity of Jetson XavierTM NX is weaker than that of the deep learning workstation, which leads to slow processing speed. In order to improve the speed, the model weight needs to be converted again, so that the inference throughput and efficiency of processing data are optimized, and the processing speed is improved. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a dam crack detection method based on an intelligent unmanned aerial vehicle, which improves the accuracy and efficiency of crack detection and can realize real-time detection by modifying YOLOv5.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is: a dam crack detection method based on an intelligent unmanned aerial vehicle, comprising the following steps: one, data set construction: using an unmanned aerial vehicle to shoot crack pictures or record videos of the dam surface under real engineering conditions, obtaining crack images by frame division of the shot videos, and obtaining a data set by manually labeling the images using labeling software;

[0007] Two, dam crack detection model construction: improving the YOLOv5 detection model to form a GE-YOLOv5 detection model, the YOLOv5 detection model is composed of four parts, the first part is the input end, the second part is the backbone extraction network, the third part is the neck network, and the fourth part is the detection end;

[0008] Three, dam crack detection model weight conversion: first, the dam surface crack data is learned and trained by the GE-YOLOv5 model detection model, and the model weight of the data set can be obtained after the training is completed, and then the GE-YOLOv5 model weight conversion is performed;

[0009] Four, edge deployment of computing processor: first, realize image burning of the computing processor on a host computer, then migrate and deploy the model processed in step three, and finally realize API calling by using a C++ program interface;

[0010] V. Dam surface crack detection of unmanned aerial vehicle: the computing processor is carried on the intelligent unmanned aerial vehicle, the intelligent unmanned aerial vehicle client is connected with the workstation on the ground through the WiFi router, the intelligent unmanned aerial vehicle takes pictures of the dam surface during the aerial photography process, and the video images taken by the camera are directly input into the computing processor after deployment for processing, and the dam surface cracks are detected by GE-YOLOv5.

[0011] In the preferred scheme, in step one, in order to prevent overfitting of the data set training, the image brightness of the data image is adjusted, the crack images obtained under different illumination intensities are simulated, and the obtained images are randomly rotated and flipped for pretreatment.

[0012] In the preferred scheme, in step two, before the feature maps of the 4th layer and the 16th layer, the 6th layer and the 12th layer, the 14th layer and the 19th layer, and the 10th layer and the 22nd layer of the YOLOv5 target detection model are fused, the 4th layer, the 6th layer, the 10th layer and the 14th layer are respectively added with a lightweight efficient channel attention mechanism ECA. First, through 1x1 convolution, the importance between different channels is learned, and the spatial feature compression of the input feature map is performed; second, the compressed feature map is learned, and the dimension of the output is still 1x1xC; finally, the channel attention is combined, and the channel attention feature Figure 1 1xC and the original input feature map HxWxC are multiplied channel by channel, and the feature map with channel attention is finally output.

[0013] In the preferred scheme, in step two, a new GhostNet module is introduced to optimize the backbone extraction network, that is, the 2nd, 4th, 6th and 8th C3 bottleneck layers in the backbone feature extraction network are replaced by GhostNet. First, the number of redundant information in the input feature map is reduced through normal 1x1 convolution to obtain a feature map with less channel information. Then, more feature maps are generated using cheap operation. Finally, the feature maps obtained by identity mapping are connected to form a new output, forming a GE-YOLOv5 detection model.

[0014] In the preferred scheme, in step three, during the initial training, the size of all data set images is restored to 640x640 pixels at the input end, and GE-YOLOv5 uses EIOU loss for training:

[0015] (1);

[0016] IOU is the intersection over union of the bounding box and the real box.

[0017] , represents the Euclidean distance of the center points of the predicted box and the real box.

[0018] The center point of the prediction box, The center point of the true bounding box;

[0019] The width of the prediction box, The width of the actual bounding box;

[0020] For the height of the predicted bounding box, The height of the actual bounding box;

[0021] C is the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box;

[0022] During training, the loss curve decreases. Training can be stopped when the loss curve is basically parallel to the X-axis. After training, the model weights for the dataset can be obtained.

[0023] In the preferred embodiment, in step three, after training is completed, the weights of the PyTorch framework model obtained from the training are converted into the TensorRT framework. The model is mainly compiled using C language, and an .engine engine file and a .dll dynamic link library are generated for deployment.

[0024] In the preferred embodiment, in step four, the computing processor is an embedded AI computing processor.

[0025] This invention provides a method for detecting cracks on dam surfaces based on intelligent unmanned aerial vehicles (UAVs). By modifying YOLOv5 to reduce computational parameters and make it more suitable for mobile deployment, multi-layer parallel dilated convolutions are added to improve feature matching capabilities for dam surface cracks. The GE-YOLOv5 model is then deployed on Jetson Xavier™ NX for edge computing, addressing the shortcomings of traditional target detection methods, such as large hardware devices, limited mobility, and inability to meet real-time requirements. Finally, a P600 UAV carrying the deployed Jetson Xavier™ NX is used to detect cracks on the dam surface, accurately determining the location of cracks in real time. This method overcomes the limitations of manual judgment and the low efficiency of current target detection algorithms, reducing manual inspection costs, improving the accuracy and efficiency of crack detection, lowering the probability of safety hazards, and achieving real-time detection and exploration of cracks on the dam surface. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0027] Figure 1 is a schematic diagram of the GhostNet module structure;

[0028] Figure 2 A diagram illustrating the lightweight and efficient channel attention mechanism, ECA.

[0029] Figure 3 This is a framework diagram of the GE-YOLOv5 model;

[0030] Figure 4 A comparison chart of the GE-YOLOv5 model before and after weight transformation;

[0031] Figure 5 Flowchart for intelligent drone-based dam surface crack detection;

[0032] Figure 6 Crack detection results for the YOLOv5 model;

[0033] Figure 7 Crack detection results for the GE-YOLOv5 model;

[0034] Figure 8 Loss curve for training the GE-YOLOv5 model; Detailed Implementation

[0035] like Figure 1 As shown, a method for detecting cracks in dams based on intelligent unmanned aerial vehicles (UAVs) includes the following steps:

[0036] I. Dataset Construction: Deep learning algorithms require a large amount of data to support training. The quality of the data has a decisive impact on model training, especially on the richness of the data samples and the accuracy of the data annotation.

[0037] Drones are used to photograph or record videos of cracks on the dam surface under actual engineering conditions. The drones sequentially pass through key points on the dam where cracks are prone to form, and finally return the collected inspection data to the starting point, ending the inspection mission. The captured video is then segmented frame to obtain crack images. Additionally, publicly available crack images can be collected online.

[0038] To prevent overfitting during dataset training, image brightness was adjusted to mimic crack images captured under varying lighting intensities, enriching the image background. Since not every drone-captured image could be taken from the same angle, the resulting images were preprocessed using Python scripts, performing random rotations and flips at different angles. This resulted in a sufficient number of images with rich backgrounds, enhancing the model's robustness. Finally, the images were manually annotated using Labelimg software, yielding a high-quality dataset.

[0039] II. Construction of Dam Crack Detection Model: The YOLOv5 detection model was improved to form the GE-YOLOv5 detection model.

[0040] YOLOv5 is an efficient and fast single-stage object detection algorithm. Therefore, based on the YOLOv5 model, it mainly consists of four parts. The first part is the input end, which mainly includes image input and Mosaic data augmentation. The second part is the backbone extraction network, which consists of convolutional layers (conv), bottleneck layers (C3), and Spatial Pyramid Pooling-Fast (SPPF). The third part is the neck network, which uses the Path Aggregation network (PANet) structure under the instance segmentation framework to enhance the network features. The last part is the detection end (YOLO Head), which decodes and predicts the obtained feature maps. Although the YOLOv5 object detection model has been applied in many fields, if it is directly applied to the surface crack detection of dam surfaces, it still faces the following problems: (1) Dam cracks appear in the form of thin, irregular black line shadows, which are difficult to identify by conventional models and have low accuracy. (2) The cracks in the dam are very similar to the surrounding environment, resulting in low accuracy and easy missed detections. (3) Currently, the video images of the dam surface are captured by drones and then imported into a deep learning workstation for processing, which results in low detection efficiency. (4) The model weights obtained after general model training are too large and cannot be well embedded in the edge AI computing processor Jetson Xavier™ NX. (5) The computing power of Jetson Xavier™ NX is weaker than that of the deep learning workstation, resulting in slow processing speed. In order to improve the speed, the model weights need to be transformed to optimize the inference throughput and efficiency of the processed data, thereby improving the processing speed.

[0041] Before fusing feature maps from layers 4 and 16, 6 and 12, 14 and 19, and 10 and 22 in the YOLOv5 object detection model, a lightweight and efficient channel attention mechanism (ECA) is added to layers 4, 6, 10, and 14 respectively. ECA is an extremely lightweight channel attention module that primarily proposes a non-dimensionality-reducing local cross-channel interaction strategy and an adaptive method for selecting the size of the one-dimensional convolutional kernel. Figure 2 As shown, firstly, 1×1 convolution is used to learn the importance of different channels and to compress the spatial features of the input feature map; secondly, channel feature learning is performed on the compressed feature map, at which point the output dimension is still 1×1×C; finally, channel attention is combined to integrate the features of the channel attention. Figure 1The original input feature map H×W×C is multiplied channel by channel to output a feature map with channel attention. This improves detection accuracy without increasing the model's computational parameters, thereby enhancing the model's ability to extract features of cracks on the dam surface and effectively improving the model's accuracy and precision in detecting cracks on the dam surface.

[0042] Large model parameters result in a large memory footprint for the trained model's recognition weights, which is not conducive to embedding in the Jetson Xavier™ NX edge AI computing processor. Furthermore, current object detection networks contain a significant amount of redundant information in the feature maps obtained from convolutional images, increasing computational costs. To remove and reduce redundancy and lower model parameters, a new GhostNet module is introduced to optimize the backbone extraction network. Specifically, GhostNet replaces the bottleneck layers C3 (layers 2, 4, 6, and 8) in the backbone feature extraction network. GhostNet is shown below. Figure 1 As shown in the figure, Φ represents a cheap linear operation, and identity represents an identity mapping.

[0043] The Ghost module first reduces the amount of redundant information in the input feature map through normal 1x1 convolution, resulting in a feature map with less channel information. Then, it uses inexpensive computation to generate more feature maps. Finally, it connects the feature maps obtained by identity mapping to form a new output, thus forming the GE-YOLOv5 detection model.

[0044] GhostNet generates more features with fewer parameters and reduces redundant information in feature maps, compressing the network while maintaining accuracy. This reduces the storage space required for the model, addresses the issue of excessively large model weights after training, reduces computational parameters, and speeds up image processing, making it more suitable for mobile deployment. The specific model framework of GE-YOLOv5 is as follows: Figure 3 As shown.

[0045] III. Weight Transformation of Dam Crack Detection Model:

[0046] First, the surface crack data of the dam is trained using the GE-YOLOv5 model. After training, the model weights of the dataset are obtained, and then the GE-YOLOv5 model weights are transformed.

[0047] Model Training: The dam surface crack data was trained using the GE-YOLOv5 model described above. At the beginning of training, all dataset images were resized to 640×640 pixels at the input. GE-YOLOv5 was trained using EIOU loss. EIOU loss considers the overlap of bounding boxes, center distance, and aspect ratio scale information on top of other loss functions, making the bounding box regression more stable and preventing divergence issues during training, unlike IOU and GIOU loss functions.

[0048] GE-YOLOv5 is trained using EIOU loss:

[0049] (1);

[0050] IOU is the intersection-union comparison between the bounding box and the ground truth box.

[0051] , representing the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box;

[0052] The center point of the prediction box, The center point of the true bounding box;

[0053] The width of the prediction box, The width of the actual bounding box;

[0054] For the height of the predicted bounding box, The height of the actual bounding box;

[0055] C is the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box;

[0056] The loss curve decreases during training, such as... Figure 8 As shown, training can be stopped when the loss curve is basically parallel to the X-axis. After training, the model weights for the dataset can be obtained.

[0057] GE-YOLOv5 model weight transformation: The weights of the PyTorch framework model trained above are transformed into those of the TensorRT framework, enabling fast and efficient inference on the GPU. The model is primarily compiled using C language, generating .engine files and .dll dynamic link libraries for deployment. This involves horizontal and vertical integration of the network layers to achieve optimization and acceleration. The optimization process does not change the number of model parameters or the underlying network computations; it only reconstructs the computation graph. While maintaining accuracy, it achieves the optimal data layer, significantly improving the processing speed of dam crack images on Jetson Xavier™ NX.

[0058] like Figure 4 (a) shows the PyTorch framework structure before GE-YOLOv5 weight transformation, and (b) shows the framework structure after weight transformation to TensorRT. As can be seen from the figures, after model weight transformation, unused output layers are eliminated to avoid unnecessary computation; on the other hand, convolution, bias, and ReLU layers are merged into a single layer, mainly through vertical and horizontal layer fusion, reducing computational steps and transmission time to minimize latency and maximize throughput in the optimized network. IV. Edge Deployment of the Computing Processor: First, the computing processor image is burned onto a host machine. An embedded AI computing processor, specifically Jetson Xavier™ NX, is selected. An environment such as Ubuntu 16.04 is installed on Jetson Xavier™ NX, supporting modules such as CUDA and CUDN. Then, the model processed in step three is migrated and deployed. Finally, API calls are implemented using a C++ program interface. Jetson Xavier™ NX provides excellent performance and energy efficiency for NVIDIA Jetson™ systems, improving model running speed while consuming less power. Deploying the proposed GE-YOLOv5 model on it can save time spent on UAV detection of dam surface cracks and reduce the cost of purchasing deep learning workstations, thereby improving detection efficiency.

[0059] V. Drone Surface Crack Detection of Dams: The intelligent drone selected is the Prometheus 600. The P600 is a medium-sized drone with strong payload capacity, long endurance, and scalability. It can be well equipped with a Jetson Xavier™ NX computing board, LiDAR, a three-axis optoelectronic pod camera (10x optical zoom), RTK, and other intelligent equipment to achieve dam surface crack detection, LiDAR obstacle avoidance, and drone position and speed pointing flight capabilities. In addition, the drone also includes: ① a 6s-10000mAh lithium polymer battery for powering the drone; ② a ubox flight controller for remote control; ③ an m8n GPS module for positioning; ④ two pairs of 15-inch folding propellers; ⑤ an S1 LiDAR, etc., which together form the intelligent drone.

[0060] The Jetson Xavier™ NX is mounted on the P600 drone and connected to a three-axis optical pod (10x optical zoom) camera. The Jetson Xavier™ NX can be powered by (1) an independent battery; (2) a connecting cable; or (3) a shared battery with the drone. Here, the Jetson Xavier™ NX is powered by connecting the drone's battery, which can reduce the overall weight of the drone and improve its stability.

[0061] The computing processor is mounted on the intelligent drone. The intelligent drone client is connected to the workstation on the ground via a WiFi router. The entire process of aerial photography by the intelligent drone can be divided into: (1) Information extraction: the three-axis photoelectric pod camera takes pictures of the dam surface; (2) Information processing: the video images captured by the three-axis photoelectric pod camera are directly input into the deployed Jetson Xavier™ NX AI calculator for processing, and the cracks on the dam surface are detected by GE-YOLOv5; (3) Information transmission: the images and videos that are detected and processed at the same time are shared to the screen of the workstation on the ground via the local area network, so that they can be observed at any time; (4) Information storage: the information of the cracks on the dam surface in each area is stored, that is, the location, coordinates, crack confidence, date and time of the cracks on the dam surface are stored in the text in real time, so as to realize intelligent identification and positioning of the cracks on the dam surface in the current monitoring screen. The process is as follows. Figure 5 .

[0062] The crack detection results identified using this detection method with the YOLOv5 model and the GE-YOLOv5 model are as follows: Figure 6 and Figure 7 As shown, the detected crack images on the dam surface are quite complex, containing hidden and micro-cracks, thus posing a certain challenge to the detection. From Figure 6 It can be observed that the YOLOv5 baseline model can roughly detect cracks on the dam surface (the cracks are shown in the boxes), but the localization is incomplete. It has low accuracy in extracting complex and fine cracks, resulting in missed detections and location errors. Figure 7 It can be seen that the improved YOLOv5 model yields the highest accuracy in dam surface crack detection, with precise localization and strong detection performance for cracks of different scales. This indicates that the improved GE-YOLOv5 detection model can improve accuracy, reduce missed detections, and is more suitable for dam crack detection. It can be used as a method or alternative for dam crack detection, thereby improving detection efficiency.

[0063] Table 1 shows the memory usage and inference speed evaluation of the YOLOv5 and GE-YOLOv5 models. As can be seen from Table 1, the GE-YOLOv5 model has a 2.7 percentage point improvement in accuracy compared to the YOLOv5 model, and its weight size is 16.3 percentage points smaller. Its inference speed is 12.1 percentage points faster than the baseline YOLOv5 model. While ensuring detection accuracy, it reduces model weight and increases inference time, thereby improving portability and enabling better deployment on Jetson Xavier™ NX.

[0064]

[0065] This detection method modifies YOLOv5, reducing computational parameters and making it more suitable for mobile deployment. It adds multi-layer parallel dilated convolutions to improve feature matching capabilities for dam surface cracks. The GE-YOLOv5 model is then deployed on Jetson Xavier™ NX for edge computing, addressing the shortcomings of traditional target detection methods, such as large hardware, limited mobility, and inability to meet real-time requirements. Finally, a P600 UAV carrying the deployed Jetson Xavier™ NX is used for dam surface crack detection, accurately determining crack locations in real time. This overcomes the limitations of manual judgment and the low efficiency of current target detection algorithms, reducing manual inspection costs, improving crack detection accuracy and efficiency, lowering the probability of safety hazards, and achieving crack detection and exploration on the dam surface.

Claims

1. A method for detecting cracks in dams based on intelligent unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: I. Dataset Construction: Using drones to capture images or record videos of cracks on the dam surface under actual engineering conditions, the captured videos are segmented into frames to obtain crack images, and the images are manually annotated using annotation software to obtain the dataset; II. Construction of Dam Crack Detection Model: The YOLOv5 detection model was improved to form the GE-YOLOv5 detection model. The YOLOv5 detection model consists of four parts: the first part is the input end, the second part is the backbone extraction network, the third part is the neck network, and the fourth part is the detection end. Before fusing the feature maps of layers 4 and 16, 6 and 12, 14 and 19, and 10 and 22 in the YOLOv5 detection model, a lightweight and efficient channel attention mechanism (ECA) is added to layers 4, 6, 10, and 14 respectively. First, 1×1 convolutions are used to learn the importance between different channels and spatial feature compression is performed on the input feature map. Second, channel feature learning is performed on the compressed feature map, and the output dimension is still 1×1×C. Finally, channel attention is combined by multiplying the channel-attention feature map (1×1×C) and the original input feature map (H×W×C) channel by channel to finally output a feature map with channel attention. A new GhostNet module is introduced to optimize the backbone extraction network. Specifically, GhostNet is used to replace the bottleneck layers C3 (layers 2, 4, 6, and 8) in the backbone feature extraction network. First, the redundant information in the input feature map is reduced by normal 1x1 convolution, resulting in a feature map with less channel information. Then, more feature maps are generated using inexpensive computation. Finally, the feature maps obtained by identity mapping are connected to form a new output, thus forming the GE-YOLOv5 detection model. III. Weight transformation of dam crack detection model: First, the dam surface crack data is trained using the GE-YOLOv5 model. After training, the model weights of the dataset can be obtained, and then the GE-YOLOv5 model weights are transformed. IV. Edge Deployment of Computing Processors: First, the image of the computing processor is burned onto a host machine. Then, the model processed in step three is migrated and deployed. Finally, API calls are implemented using a C++ program interface. V. Detection of dam surface cracks by drone: The computing processor is mounted on a smart drone. The smart drone client is connected to the workstation on the ground via a WiFi router. During the aerial photography process, the smart drone takes pictures of the dam surface. The video images captured by the camera are directly input into the deployed computing processor for processing. Cracks on the dam surface are detected by GE-YOLOv5.

2. The method for detecting dam cracks based on intelligent unmanned aerial vehicles according to claim 1, characterized in that, In step one, to prevent overfitting of the dataset during training, the brightness of the data images is adjusted to mimic crack images obtained under different light intensities. The resulting images are then preprocessed by randomly rotating and flipping them at different angles.

3. The method for detecting dam cracks based on intelligent unmanned aerial vehicles according to claim 1, characterized in that, In step three, at the beginning of training, all dataset images are resized to 640×640 pixels at the input end, and GE-YOLOv5 is trained using EIOU loss. (1)； IOU is the intersection-union comparison between the bounding box and the ground truth box. , representing the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; The center point of the prediction box, The center point of the true bounding box; The width of the prediction box, The width of the actual bounding box; For the height of the predicted bounding box, The height of the actual bounding box; C is the diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; During training, the loss curve decreases. Training can be stopped when the loss curve is basically parallel to the X-axis. After training, the model weights for the dataset can be obtained.

4. The method for detecting dam cracks based on intelligent unmanned aerial vehicles according to claim 1, characterized in that, In step three, after training is completed, the weights of the PyTorch framework model obtained from the training are converted into the TensorRT framework. The model is mainly compiled using C language, and .engine files and .dll dynamic link libraries are generated for deployment.

5. The method for detecting dam cracks based on intelligent unmanned aerial vehicles according to claim 1, characterized in that, In step four, an embedded AI computing processor is selected as the computing processor.

Images