Dam apparent structure defect unmanned aerial vehicle aerial photography inspection method and system

Abstract

The invention provides an unmanned aerial vehicle aerial photography inspection method and system for dam apparent structure defects, and relates to the field of hydraulic engineering and computer vision. The inspection method comprises the following steps: performing low-altitude aerial photography by using an unmanned aerial vehicle to obtain a dam apparent structure defect high-definition image; performing data enhancement processing on the dam apparent structure defect high-definition image based on a generative adversarial network to obtain dam apparent structure defect image data, and performing image label labeling based on Label; on the basis of a convolutional attention mechanism (CBAM) and a SIOU loss function, optimizing a YOLOv8 target detection network to establish a CS-YOLOv8 network, and training the CS-YOLOv8 network on the basis of the marked dam apparent structure defect image data; and integrating the trained CS-YOLOv8 network and slice-assisted super reasoning, implanting the integrated CS-YOLOv8 network and slice-assisted super reasoning into an unmanned aerial vehicle system, and intelligently inspecting dam apparent structure defects in real time based on unmanned aerial vehicle aerial photography. According to the inspection method provided by the invention, the dam can be inspected in real time to perceive the apparent defects, and the detection precision of the dam apparent small target hidden structure defects is improved.

Description

Technical Field

[0001] The present invention relates to the fields of water conservancy engineering and computer vision, and in particular to a method and system for unmanned aerial vehicle (UAV) aerial photography inspection of apparent structural defects of dams. Background Art

[0002] Earth-rock dams are an important part of the flood control engineering system and a reliable barrier to resist flood threats and ensure flood control safety. However, due to the constraints of exploration, design, and construction technology, there are widespread hidden dangers and diseases such as uneven soil, cracks, voids, and loose bodies inside the dams. Under complex water conservancy conditions such as floods, accidents such as dam pipe bursts and breaches caused by hidden dangers and diseases occur frequently. Apparent defects such as cracks and pits in dams are the external manifestations of hidden dangers and diseases inside dams. Timely perception of apparent defects in dams is the key to the safety management of dam projects. At present, the perception of apparent defects in dams mainly relies on manual inspection methods. For earth-rock dams with long lines and wide areas, there are problems such as time-consuming and labor-intensive, high missed detection rate, and high operation risks, which makes it difficult to meet the actual needs of dam emergency rescue.

[0003] In recent years, thanks to the rapid development of technologies such as drone aerial photography and computer vision, many scholars have organically combined drones and computer vision technologies and successfully applied them to the non-destructive testing of surface defects of engineering structures. With the advantages of drones such as flexibility and portability, simple operation, high cruising efficiency and strong adaptability to complex environments, the drones can remotely and quickly shoot videos or images of engineering surface defects through the imaging equipment they carry, and then combine computer vision technology based on deep learning algorithms to extract image features such as shape, texture and color of engineering surface defects, thereby identifying the type and location information of surface defects. Therefore, how to combine drones with computer vision is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for drone aerial photography inspection of dam surface structural defects, so as to solve the problem that the inspection of dam structure defects is time-consuming and labor-intensive.

[0005] The present invention provides a method for drone aerial photography inspection of dam surface structural defects, which adopts the following technical solutions: Use drones to take low-altitude aerial photos to collect apparent images of the dam structure, and obtain high-definition images of the dam's apparent structural defects through cropping and screening; Based on the generative adversarial network, data enhancement processing is performed on the high-definition images of the apparent structural defects of the dam to obtain the image data of the apparent structural defects of the dam, and image labels are annotated based on Labelme; Based on the convolutional attention mechanism CBAM and SIoU loss function, the YOLOv8 target detection network is optimized to establish the CS-YOLOv8 network, and the CS-YOLOv8 network is trained with the labeled dam surface structure defect image data; The trained CS-YOLOv8 network is integrated with slice-assisted hyper-inference and implanted into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

[0006] Optionally, a drone is controlled to fly along the dam for aerial photography at different altitudes and angles to obtain high-definition image data of the dam's surface, including image data of apparent cracks, pits and defects on the dam's surface.

[0007] Optionally, data enhancement processing is performed on the high-definition images of dam surface structural defects based on a generative adversarial network to obtain the image data of the dam surface structural defects, including: using a generative adversarial network to perform data enhancement processing on the high-definition images of dam surface structural defects to generate a large number of scene-realistic virtual image samples of dam surface structural defects, and establishing the image data of dam surface structural defects based on the virtual image samples.

[0008] Optionally, the process of labeling images based on Labelme includes: using Labelme to manually label the location, type and range of defects in the image data of the dam surface structure defects, and generating a corresponding label file for each image.

[0009] Optionally, the process of optimizing the YOLOv8 object detection network to build a CS-YOLOv8 network based on the convolutional attention mechanism CBAM and SIoU loss function includes: The convolutional attention mechanism CBAM is embedded into the backbone network of the YOLOv8 target detection network to enhance the feature extraction ability of the YOLOv8 target detection network. At the same time, the positioning loss function in the YOLOv8 target detection network is replaced by the SIoU loss function, thereby establishing a CS-YOLOv8 network with improved image feature extraction capabilities and small target structure defect recognition performance.

[0010] Optionally, the process of embedding the convolutional attention mechanism CBAM into the backbone network of the YOLOv8 target detection network includes: setting hyperparameters of the YOLOv8 target detection network, integrating the convolutional attention mechanism CBAM into four C2f module positions of the backbone network respectively, and performing ablation experiments using the dam apparent defect dataset to evaluate the C2f module position where the convolutional attention mechanism CBAM is best inserted.

[0011] Optionally, in the process of setting the hyperparameters of the YOLOv8 target detection network, the hyperparameters include training rounds, initial learning rate, weight decay rate, batch size, image size, and the indicators for evaluating the optimal insertion position of the C2f module of the convolutional attention mechanism CBAM include precision, recall, F1 score, average precision, number of detection frames, and weight file size.

[0012] Optionally, the process of integrating the trained CS-YOLOv8 network with slice-assisted hyper-inference includes: adding a slice-assisted hyper-inference module to the optimized CS-YOLOv8 network detection end and integrating it into the system to obtain an optimal slice size.

[0013] Optionally, the process of obtaining the optimal slice size includes: slicing a single original image into different sizes by means of an auxiliary super-inference sliding window to obtain sub-images of slices of each size, and reasoning on each sub-image and the original image respectively to complete the extraction and perception of small target features, and then using a non-maximum suppression algorithm to combine the full reasoning of the original image with the prediction results after slice-assisted super-inference, and finally obtaining prediction results of different slice sizes of a single image, and evaluating the optimal slice size based on the prediction results.

[0014] In a second aspect of the present invention, a drone aerial photography inspection system for apparent structural defects of dams is provided, comprising a data acquisition module, a data processing module, a network optimization training module and an integration module, wherein: Data acquisition module: configured to collect the surface image of the dam structure by using low-altitude aerial photography of unmanned aerial vehicles, and to screen the high-definition images of the dam surface structure defects by cropping; Data processing module: configured to perform data enhancement processing on the high-definition image of the apparent structure defect of the dam based on the generative adversarial network, obtain the image data of the apparent structure defect of the dam, and annotate the image labels based on Labelme; Network optimization training module: configured to optimize the YOLOv8 target detection network based on the convolutional attention mechanism CBAM and SIoU loss function to establish a CS-YOLOv8 network, and train the CS-YOLOv8 network in combination with the labeled dam surface structure defect image data; Integration module: Integrate the trained CS-YOLOv8 network with slice-assisted hyper-inference and implant it into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

[0015] The invention provides a method and system for drone aerial photography inspection of dam surface structural defects, which has the following beneficial effects: 1. The present invention integrates the trained CS-YOLOv8 network with slice-assisted hyper-inference and implants it into the UAV system, thereby realizing the real-time perception of apparent defects in the dam during UAV inspection, thereby increasing safety, efficiency and speed; 2. The present invention improves the detection effect of small targets in drone aerial images by fusing slices to assist hyper-inference. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method for drone aerial photography inspection of dam surface structural defects provided by the present invention; Figure 2 A schematic diagram for comparing the average accuracy of the models provided by the present invention; Figure 3 A schematic diagram of the actual perceived effect of dam defects provided by the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0018] The embodiment of the present invention provides a method for drone aerial photography inspection of dam surface structural defects, see Figure 1 ,include: S1. Use drones to take low-altitude aerial photos to collect apparent images of the dam structure, and obtain high-definition images of the apparent structural defects of the dam through cropping and screening; S2. Perform data enhancement processing on the high-definition image of the apparent structure defect of the dam based on the generative adversarial network to obtain the image data of the apparent structure defect of the dam, and annotate the image labels based on Labelme; S3. Based on the convolutional attention mechanism CBAM and SIoU loss function, the YOLOv8 target detection network is optimized to establish the CS-YOLOv8 network, and the CS-YOLOv8 network is trained with the labeled dam surface structure defect image data; S4. Integrate the trained CS-YOLOv8 network with slice-assisted hyper-inference and implant it into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

[0019] In some embodiments, the process of executing step S1 includes: S1.1. Use drones to take low-altitude aerial photos to collect surface images of dam structures; S1.2. Obtain high-definition images of dam surface structural defects through cropping and screening.

[0020] Specifically, when executing step S1.1, using a drone to take low-altitude aerial photos to collect the surface image of the dam structure, it includes: controlling the drone to fly along the dam for aerial photography, using different altitudes and angles for aerial photography, and obtaining high-definition image data of the dam surface, wherein the high-definition image data includes image data of apparent cracks, pits and defects on the dam surface.

[0021] In fact, when controlling the drone to fly along the embankment for aerial photography, the aerial photography height is 10-15m, and the aerial photography angles include looking down and oblique.

[0022] Specifically, when executing step S1.2, obtaining a high-definition image of the dam's apparent structural defects by cropping and screening, it includes: screening the high-definition image data, eliminating incomplete or undistinguishable image data of the dam's apparent cracks and pits, selecting images with apparent defects, and cropping the images to a uniform specification.

[0023] In some embodiments, the process of executing step S2 includes: S2.1. Based on the generative adversarial network, data enhancement processing is performed on the high-definition image of the apparent structure defect of the dam to obtain the image data of the apparent structure defect of the dam; S2.2. Image labeling based on Labelme.

[0024] Specifically, when executing step S2.1, a generative adversarial network is used to perform data enhancement processing on the high-definition image of the dam's apparent structural defects to generate a large number of scene-realistic virtual image samples of the dam's apparent structural defects, and the dam's apparent structural defect image data is established based on the virtual image samples, and the generative adversarial network includes CycleGAN.

[0025] In fact, using a generative adversarial network to perform data enhancement processing on high-definition images of dam surface structural defects can achieve style transfer of dam surface defect images under various weather conditions, imaging angles and multiple backgrounds, so as to generate a large number of dam surface defect training samples to help enhance the generalization ability of the model.

[0026] Specifically, when executing step S2.2, it includes: using Labelme to manually annotate the location, range and type of defects in the dam surface structure defect image data, and generating a corresponding label file for each image, thereby constructing a dam surface defect image data set.

[0027] In fact, in the process of constructing the dam surface defect dataset, the images in the dataset will be randomly divided into training sets and validation sets in a certain proportion for subsequent training and optimization of the target detection model.

[0028] In some embodiments, the process of executing step S3 includes: S3.1. Based on the convolutional attention mechanism CBAM and SIoU loss function, the YOLOv8 target detection network is optimized to establish the CS-YOLOv8 network. S3.2. Train the CS-YOLOv8 network based on the labeled dam surface structural defect image data.

[0029] Specifically, when executing step S3.1, based on the convolutional attention mechanism CBAM and the SIoU loss function, the YOLOv8 target detection network is optimized to establish a CS-YOLOv8 network, including: embedding the convolutional attention mechanism CBAM into the YOLOv8 target detection network backbone network to enhance the feature extraction capability of the YOLOv8 target detection network, and replacing the positioning loss function in the YOLOv8 target detection network with the SIoU loss function, thereby establishing a CS-YOLOv8 network with improved image feature extraction capability and small target structure defect recognition performance.

[0030] In fact, the convolutional attention mechanism CBAM (Convolutional Block Attention Module, CBAM) consists of a channel attention module and a spatial attention module in series. In the channel attention module, firstly, the maximum pooling and average pooling operations are performed on each channel of the input feature map F1 to compress its spatial dimension; secondly, the pooling result is calculated by the fully connected layer, and the two calculation results are added and the weight of each channel of the input feature map is obtained after the Sigmoid activation function; finally, the weight is multiplied by the original input feature map to obtain the channel attention feature map F2. In the spatial attention module, firstly, the maximum pooling and average pooling operations are performed on the channel attention feature map F2 along the channel dimension; secondly, the features after the maximum pooling and average pooling are spliced ​​along the channel dimension, and this feature map is processed by the convolution layer to generate the original representation of the spatial attention weight. Then, the Sigmoid activation function is applied to normalize these weights to obtain the spatial attention weights that can be directly used for weighted feature maps. Finally, the spatial attention weights are multiplied by the original input feature map F1 to obtain the final spatial attention feature map F3.

[0031] Specifically, when executing step S3.2, training the CS-YOLOv8 network in combination with the labeled dam surface structural defect image data, it includes: setting the hyperparameters of the YOLOv8 target detection network, integrating the convolutional attention mechanism CBAM into the four C2f module positions of the backbone network respectively, and performing ablation experiments using the dam surface defect dataset to evaluate the C2f module position where the convolutional attention mechanism CBAM is best inserted.

[0032] Furthermore, in the process of setting the hyperparameters of the YOLOv8 target detection network, the hyperparameters include training rounds, initial learning rate, weight decay rate, batch size, and image size. The indicators for evaluating the optimal insertion position of the C2f module of the convolutional attention mechanism CBAM include precision, recall, F1 score, average precision, number of detection frames, and weight file size.

[0033] In fact, the present invention integrates the CBAM module with the C2f module. Considering that the second to fourth C2f module layers are the structure of outputting feature maps to the neck, the CBAM module is first inserted into the second to fourth C2f modules, and training experiments are carried out. When the CBAM module is inserted into the fourth C2f module, the precision, recall and average precision of the model are improved.

[0034] Furthermore, considering that the CBAM module can be inserted into multiple C2f modules, the second and third C2f modules can be inserted into the CBAM module in addition to the fourth C2f module to try to further optimize the network performance. The best fusion effect of the CBAM module alone and the fourth C2f module of the YOLOv8 backbone network is obtained through hyperparameter evaluation.

[0035] For further information, see Figure 2 , based on finding the best insertion position of the CBAM module, the SIoU loss function is used to replace the original positioning loss function of the YOLOv8 target detection network. The experiment uses the target detection network YOLOv8 as the benchmark network, and each module is added to the network alone or in combination for training to verify whether the optimized CS-YOLOv8 network model is the best model. In the experiment, except for the added modules as variables, the other control factors remain unchanged to ensure the reliability of the experiment. The 100-500 rounds of training convergence process of the model under the four situations, among which the average accuracy of the optimized CS-YOLOv8 network model begins to stabilize after 350 rounds, and both mAP@0.5 and mAP@0.5:0.95 show excellent performance. mAP@0.5 represents the mAP calculated when the IoU (Intersection over Union) threshold is 0.5, and mAP@0.5:0.95 represents the average mAP under multiple IoU thresholds from 0.5 to 0.95.

[0036] In some embodiments, the process of executing step S4 includes: S4.1. Integrate the trained CS-YOLOv8 network with slice-assisted hyper-inference. S4.2. Implant a drone system to conduct real-time intelligent inspection of the dam's apparent structural defects based on drone aerial photography.

[0037] Specifically, when executing step S4.1, integrating the trained CS-YOLOv8 network with slice-assisted hyper-inference, it includes: adding a slice-assisted hyper-inference module to the optimized CS-YOLOv8 network detection end and integrating it into the system to obtain the optimal slice size.

[0038] Furthermore, a single original image is sliced ​​into different sizes by means of an assisted hyper-inference sliding window to obtain sub-images of slices of each size, and each sub-image and the original image are reasoned separately to complete the extraction and perception of small target features. Then, the full reasoning of the original image is combined with the prediction results after slice-assisted hyper-inference using a non-maximum suppression algorithm, and finally the prediction results of different slice sizes of a single image are obtained, and the optimal slice size is evaluated based on the prediction results.

[0039] In fact, see Figure 3 When the CS-YOLOv8 network detection end adds the slice-assisted hyper-inference module (SAHI), and the slice size is 1280 pixels, the original image is cut into 30 slices. Compared with the detection of YOLOv8 network or CS-YOLOv8 network alone, whether it is 10m overhead aerial photos, 15m overhead aerial photos, style transfer photos, or oblique aerial photos, the pit defects of the dam can be well identified. However, even with the non-maximum suppression algorithm, the crack defects still have the problem of repeated detection and counting. When the CS-YOLOv8 network detection end adds the slice-assisted hyper-inference module (SAHI), and the slice size is 1920 pixels, the original image is cut into 9 slices. Compared with the detection of YOLOv8 network or CS-YOLOv8 network alone, the pit defects of the dam can be well identified. YOLOv8 network or CS-YOLOv8 network detection, whether it is a 10m overhead aerial photo, a 15m overhead aerial photo, a style transfer map, or an oblique aerial photo, the pit defects of the dam can be well identified, and only a small number of cracks are repeatedly detected by the network, showing a good detection effect; when the CS-YOLOv8 network detection end adds a slice-assisted hyper-inference module (SAHI), and the slice size is 2560 pixels, and the original image is cut into 4 slices, although the network model does not have the situation of repeated crack detection, there are certain missed detections; therefore, in the case of the aerial image of the present invention, the perception effect of the apparent defect of the dam is optimal when the slice size is 1920 pixels.

[0040] Further, execute step S4.2, implant the integrated system into the UAV system, and conduct real-time intelligent inspection of apparent structural defects of the dam based on UAV aerial photography.

[0041] The embodiment of the present invention provides a drone aerial photography inspection system for apparent structural defects of dams, including a data acquisition module, a data processing module, a network optimization training module and an integration module, wherein: Data acquisition module: configured to collect the surface image of the dam structure by using low-altitude aerial photography of unmanned aerial vehicles, and to screen the high-definition images of the dam surface structure defects by cropping; Data processing module: configured to perform data enhancement processing on the high-definition image of the apparent structure defect of the dam based on the generative adversarial network, obtain the image data of the apparent structure defect of the dam, and annotate the image labels based on Labelme; Network optimization training module: configured to optimize the YOLOv8 target detection network based on the convolutional attention mechanism CBAM and SIoU loss function to establish a CS-YOLOv8 network, and train the CS-YOLOv8 network in combination with the labeled dam surface structure defect image data; Integration module: Integrate the trained CS-YOLOv8 network with slice-assisted hyper-inference and implant it into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

[0042] Although the embodiments of the present invention are described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein may have other embodiments and may be implemented or realized in a variety of ways.

Claims

1. A method for drone aerial photography inspection of dam surface structural defects, characterized in that: The following steps are involved: Use drones to take low-altitude aerial photos to collect apparent images of the dam structure, and obtain high-definition images of the dam's apparent structural defects through cropping and screening; Based on the generative adversarial network, data enhancement processing is performed on the high-definition images of the apparent structural defects of the dam to obtain the image data of the apparent structural defects of the dam, and image labels are annotated based on Labelme; Based on the convolutional attention mechanism CBAM and SIoU loss function, the YOLOv8 target detection network is optimized to establish the CS-YOLOv8 network, and the CS-YOLOv8 network is trained with the labeled dam surface structure defect image data; The trained CS-YOLOv8 network is integrated with slice-assisted hyper-inference and implanted into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

2. The method for drone aerial photography inspection of dam surface structural defects according to claim 1 is characterized in that: The process of using drones to take low-altitude aerial photos to collect surface images of dam structures includes: The drone is controlled to fly along the dam for aerial photography at different altitudes and angles to obtain high-definition image data of the dam surface, including image data of cracks, pits and defects on the dam surface.

3. The method for drone aerial photography inspection of dam surface structural defects according to claim 1 is characterized in that: Based on the generative adversarial network, the data enhancement processing of the high-definition image of the apparent structure defect of the dam is carried out to obtain the image data of the apparent structure defect of the dam, including: A generative adversarial network is used to perform data enhancement processing on high-definition images of dam surface structural defects, so as to generate a large number of realistic virtual image samples of dam surface structural defects, and the image data of dam surface structural defects are established based on the virtual image samples.

4. The method for drone aerial photography inspection of dam surface structural defects according to claim 1 is characterized in that: The process of labeling images based on Labelme includes: Labelme was used to manually annotate the location, type and range of defects in the dam surface structural defect image data, and a corresponding label file was generated for each image.

5. The method for drone aerial photography inspection of dam surface structural defects according to claim 1 is characterized in that: Based on the convolutional attention mechanism CBAM and SIoU loss function, the process of optimizing the YOLOv8 target detection network to establish the CS-YOLOv8 network includes: The convolutional attention mechanism CBAM is embedded into the backbone network of the YOLOv8 target detection network to enhance the feature extraction ability of the YOLOv8 target detection network. At the same time, the positioning loss function in the YOLOv8 target detection network is replaced by the SIoU loss function, thereby establishing a CS-YOLOv8 network with improved image feature extraction capabilities and small target structure defect recognition performance.

6. The method for drone aerial photography inspection of dam surface structural defects according to claim 5 is characterized in that: The process of embedding the convolutional attention mechanism CBAM into the YOLOv8 target detection network backbone network includes: The hyperparameters of the YOLOv8 target detection network are set, and the convolutional attention mechanism CBAM is integrated into the four C2f module positions of the backbone network respectively. The ablation experiment is carried out using the dam apparent defect dataset to evaluate the optimal C2f module position for the convolutional attention mechanism CBAM.

7. The method for drone aerial photography inspection of dam surface structural defects according to claim 6 is characterized in that: In the process of setting the hyperparameters of the YOLOv8 target detection network, the hyperparameters include training rounds, initial learning rate, weight decay rate, batch size, and image size. The indicators for evaluating the optimal insertion position of the C2f module of the convolutional attention mechanism CBAM include precision, recall, F1 score, average precision, number of detection frames, and weight file size.

8. The method for drone aerial photography inspection of dam surface structural defects according to claim 1 is characterized in that: The process of integrating the trained CS-YOLOv8 network with slice-assisted hyper-inference includes: A slice-assisted hyper-inference module is added to the optimized CS-YOLOv8 network detection end and integrated into the system to obtain the optimal slice size.

9. The method for drone aerial photography inspection of dam surface structural defects according to claim 8 is characterized in that: The process of obtaining the optimal slice size includes: A single original image is sliced ​​into different sizes by means of an auxiliary super-inference sliding window to obtain sub-images of slices of different sizes. Inference is performed on each sub-image and the original image respectively to extract and perceive the features of small targets. Then, the full reasoning of the original image is combined with the prediction results after slice-assisted super-inference using the non-maximum suppression algorithm. Finally, the prediction results of different slice sizes of a single image are obtained, and the optimal slice size is evaluated based on the prediction results.

10. A drone aerial photography inspection system for dam surface structural defects, characterized in that: It includes data acquisition module, data processing module, network optimization training module and integration module, among which: Data acquisition module: configured to collect the surface image of the dam structure by using low-altitude aerial photography of unmanned aerial vehicles, and to screen the high-definition images of the dam surface structure defects by cropping; Data processing module: configured to perform data enhancement processing on the high-definition image of the apparent structure defect of the dam based on the generative adversarial network, obtain the image data of the apparent structure defect of the dam, and annotate the image labels based on Labelme; Network optimization training module: configured to optimize the YOLOv8 target detection network based on the convolutional attention mechanism CBAM and SIoU loss function to establish a CS-YOLOv8 network, and train the CS-YOLOv8 network in combination with the labeled dam surface structure defect image data; Integration module: Integrate the trained CS-YOLOv8 network with slice-assisted hyper-inference and implant it into the drone system to conduct real-time intelligent inspection of the apparent structural defects of the dam based on drone aerial photography.

Images