Expressway defect detection method for inspection unmanned aerial vehicle

Abstract

The invention discloses a highway defect detection method for an inspection unmanned aerial vehicle, and the method comprises the following steps: S1, employing the inspection unmanned aerial vehicle to shoot image data under a plurality of highway defect types, and carrying out the enhancement of the image data; s2, performing defect labeling on the enhanced image data; s3, constructing a highway defect detection model based on an improved YOLOv8n algorithm; s4, training an expressway defect detection model by using the image data subjected to defect labeling in the step S2 to obtain a trained expressway defect detection model; s5, after the inspection unmanned aerial vehicle is used for shooting expressway defect images, the images are sent into the trained expressway defect detection model, and defect detection is completed. By integrating unmanned aerial vehicle collection, AI intelligent analysis and big data management, automation and intelligence of highway inspection are realized, so that the efficiency and safety of inspection work are improved, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to highway defect detection, and particularly to a highway defect detection method for inspection drones. Background Art

[0002] With the rapid economic development and the acceleration of the urbanization process, highways, as an important part of the modern transportation network, the importance of their safe operation and maintenance has become increasingly prominent. The health status of highways is directly related to the public's travel safety and traffic efficiency. Therefore, regular and efficient inspection of highways is a key link to ensure their safe operation.

[0003] Traditional highway inspections mainly rely on manual inspections, and these methods have many limitations, including low inspection efficiency, limited coverage, poor safety, and high costs. Especially in the inspection of key parts such as bridges and slopes, it is difficult to achieve ideal results with manual inspections. These parts are often difficult to access due to the special geographical location, or the inspection environment is dangerous, posing a threat to the safety of inspection personnel. And the long-term employment of professional inspection personnel is costly, and requires supporting safety equipment and training.

[0004] In recent years, the development of drone technology has provided a new solution for highway inspections. Drones can enter areas that are difficult for humans to reach and conduct aerial inspections, quickly collecting a large amount of image data. However, simply relying on drones to collect data is far from enough. How to efficiently and accurately process these data and identify road defects from them is the key to improving the inspection efficiency and quality.

[0005] Traditional image processing algorithms often require complex preprocessing and parameter adjustment, and have high requirements for professional knowledge. In a complex and changing environment, the accuracy of defect recognition by traditional algorithms is relatively low. Traditional algorithms have a slow processing speed and are difficult to meet the needs of real-time monitoring. In this context, the development of artificial intelligence and big data technology provides the possibility to solve this problem. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a highway defect detection method for inspection drones. By integrating drone collection, AI intelligent analysis, and big data management, the automation and intelligence of highway inspections are realized to improve the efficiency and safety of inspection work and reduce maintenance costs.

[0007] The purpose of the present invention is achieved through the following technical solutions: A highway defect detection method for inspection drones includes the following steps

[0008] S1. Use an inspection drone to capture image data under multiple highway defect types and enhance the image data;

[0009] S2. Perform defect annotation on the enhanced image data;

[0010] S3. Build a highway defect detection model based on the improved YOLOv8n algorithm;

[0011] S4. Use the image data with defect annotation in step S2 to train the highway defect detection model to obtain a trained highway defect detection model;

[0012] S5. After using the inspection drone to take highway defect images, send the images into the trained highway defect detection model to complete defect detection.

[0013] The beneficial effects of the present invention are: (1) Improve inspection efficiency: The drone can complete the inspection tasks of large areas in a short time, avoiding the time-consuming and laborious problems of traditional manual inspection methods. By optimizing the object detection model through lightweight technology, the inspection efficiency of the drone can be further improved, realizing fast and accurate defect detection.

[0014] (2) Reduce resource consumption: The lightweight technology reduces the computational complexity and memory occupancy of the model, enabling the object detection model to run efficiently on resource-limited embedded devices. This helps to reduce the energy consumption and cost of the drone and improve the sustainability of the inspection tasks.

[0015] (3) Ensure industrial safety: Drone inspection can accurately detect potential safety hazards, providing strong support for ensuring the safe operation of industrial equipment. The application of lightweight technology will further enhance the reliability and accuracy of drone inspection, providing a more solid guarantee for industrial safety. Brief Description of the Drawings

[0016] Figure 1 is the flowchart of the method of the present invention;

[0017] Figure 2 is the principle comparison diagram of PConv and conventional Conv;

[0018] Figure 3 is the structural schematic diagram of the DWRSeg network;

[0019] Figure 4 is the overall network framework schematic diagram. Detailed Embodiments

[0020] The technical solution of the present invention will be further described in detail below with reference to the drawings, but the protection scope of the present invention is not limited to the following.

[0021] As Figure 1As shown in the figure, a highway defect detection method for inspection drones includes the following steps

[0022] S1. Use an inspection drone to capture image data under multiple highway defect types and enhance the image data;

[0023] The present invention selects the DJI M350RTK drone equipped with the Zenmuse P1 camera to collect picture data of highway slopes and bridges, based on in-depth understanding of these two products and their outstanding performance in the field of aerial photography and mapping.

[0024] The DJI M350RTK drone, as a professional mapping drone, is designed to meet the needs of high-precision and high-efficiency aerial photography and mapping. Its excellent flight stability benefits from an advanced flight control system and a powerful power system. Even in the complex and changeable highway environment, it can maintain a stable flight attitude, ensuring clear and stable captured images. At the same time, the battery life of the M350RTK reaches 55 minutes, which is leading in the same type of products, meaning it can cover a wider area in one takeoff, greatly improving the efficiency of data collection.

[0025] In terms of positioning and obstacle avoidance, the M350RTK adopts RTK (Real-Time Kinematic) technology, which can achieve centimeter-level precise positioning, which is crucial for application scenarios such as highway slopes and bridges with extremely high precision requirements. In addition, it is equipped with a six-way positioning and obstacle avoidance system, including front view, rear view, downward view, left view, right view, and upward view obstacle avoidance sensors, which can sense the surrounding environment in real time during flight, effectively avoid collisions with obstacles on the highway, and ensure the safe progress of data collection.

[0026] The Zenmuse P1 camera, on the other hand, is a high-performance aerial survey payload designed for the professional mapping field. It integrates a full-frame image sensor with a pixel count as high as 45 million, capable of capturing more delicate and real image details. At the same time, the Zenmuse P1 camera also adopts a mechanical global shutter, effectively avoiding image distortion caused by motion blur and ensuring the quality of the captured images. In addition, it supports a variety of fixed-focus lenses, capable of meeting the shooting needs in different scenarios. Whether it is details such as cracks and spalls on slopes or features such as structures and deformations of bridges, they can be clearly and accurately recorded.

[0027] In terms of data processing, the Zenmuse P1 camera also performs excellently. It supports the intelligent swing shooting function, which can automatically adjust the shooting angle and focal length according to the preset flight path and shooting parameters, achieving large-area and high-efficiency data collection. At the same time, the Zenmuse P1 camera also has powerful data processing capabilities, capable of preprocessing and correcting images in real time during shooting, ensuring the accuracy and reliability of the data.

[0028] In summary, the combination of the DJI M350RTK drone and the Zenmuse P1 camera, with its excellent flight stability, high-precision positioning and obstacle avoidance system, high-performance optical sensor, and powerful data processing capabilities, has become the preferred solution for collecting highway slope protection and bridge picture data in this project. This combination can not only improve the efficiency and accuracy of data collection but also provide strong support for subsequent disease identification and analysis, ensuring the safety maintenance of highways.

[0029] Use data augmentation techniques such as rotation, flipping, brightness enhancement, contrast enhancement, and noise introduction to improve the generalization ability and robustness of the model;

[0030] S2. Label the defects in the augmented image data;

[0031] S3. Build a highway defect detection model based on the improved YOLOv8n algorithm;

[0032] Facing the problems of few pixels occupied by targets, dense distribution, and class imbalance in the highway inspection drone dataset, the original YOLOv8n has problems such as missed detection and false detection and low accuracy. In view of the characteristics of the highway inspection drone dataset, the following improvements are made:

[0033] (1) EIoU loss function:

[0034] The traditional CIoU loss function has a problem that it optimizes the width-to-height ratio difference between the predicted box and the ground truth box as a whole, which may lead to an unreasonable optimization direction. To solve this problem, this paper introduces the EIoU loss function, which decomposes the width-to-height ratio difference into two independent parts, width and height, for optimization, thus more effectively improving the positioning accuracy.

[0035] The original YOLOv8n model uses CIoU as the bounding box regression function to calculate the similarity between bounding boxes in object recognition. CIoU is an improvement over the traditional intersection over union (IoU), which is used to measure the overlap between the predicted box and the ground truth box. However, IoU only considers the position information of the boxes and ignores the size and aspect ratio of the boxes, resulting in an inability to accurately reflect the similarity between two boxes in some cases. In contrast, CIoU incorporates the center point distance, aspect ratio difference, and box size into its calculation. The CIoU loss function is defined as follows:

[0036]

[0037] where ρ(·) = ∥b - b gt ∥ 2 denotes the distance between b and b gtThe Euclidean distance between them, where c represents the diagonal length of the smallest bounding box containing these two boxes. The gradient expressions of v with respect to w and h are as follows:

[0038]

[0039] Since v only reflects the aspect ratio difference rather than the true relationship between the predicted box and the ground truth box, CIoU may optimize the similarity in an unreasonable way. To solve this problem, we further improve CIoU to a more effective EIoU.

[0040] L EIoU = L IoU + L dir + L asp (6)

[0041]

[0042] To optimize the bounding box regression more effectively, we decompose the loss function into three parts. The Intersection over Union loss L IoU : measures the overlap degree between the predicted box and the ground truth box; the distance loss L dir : considers the distance between the centers of the predicted box and the ground truth box; the aspect ratio loss L asp : focuses on the difference in aspect ratios between the predicted box and the ground truth box. Among them, the height-width loss directly minimizes the differences in height and width between the predicted box and the ground truth box, thereby accelerating model convergence and improving localization accuracy.

[0043] Experimental results show that the LSI-YOLOv8 model using the EIoU loss function outperforms the model using the CIoU loss function in terms of metrics such as mAP@0.5 and mAP@0.5:0.95.

[0044] (2) PC2f module:

[0045] The C2f module in YOLOv8 is used to fuse feature maps of different scales, but it also increases the computational complexity and the number of parameters of the model. To solve this problem, this paper introduces PC2f, which is formed by fusing the partial convolution PConv module with C2f. That is, conventional Conv is only applied for spatial feature extraction on a part of the input channels, while the remaining channels remain unchanged. The comparison diagram of the principle of PConv

[0046] and conventional Conv is as follows Figure 2 shown.

[0047] (3) Attention mechanism:

[0048] The attention mechanism aims to highlight key information by assigning higher weights to specific regions. This promotes the fusion and interaction of information between different levels or scales and selectively enhances the feature channels rich in target information. Ultimately, this enables the model to prioritize important regions and minimize the influence of background information.

[0049] In addition, the present system introduces an attention mechanism in the C2f module to improve the model's feature selection ability and make it more focused on the features of bridges and slope protection defects on highways. The present system uses the DWRSeg network to introduce the attention mechanism, which combines a novel dilated-wise residual (DWR) module and a simple inverted residual (SIR) module for high and low levels.

[0050] The module structure is as follows Figure 3 shown;

[0051] The DWRSeg network of the present invention adopts an encoder-decoder architecture. The encoder part consists of four main stages:

[0052] The initial stem cell block, the primary stage of the SIR module, and two advanced stages of the DWR module. The DWR module adopts a residual structure and effectively collects multi-scale context information through a two-stage strategy within the residual, achieving the fusion of feature maps with multi-scale receptive fields. This method splits the original single-stage multi-scale context information acquisition process into two stages, simplifying the information acquisition process. The SIR module, derived from the DWR module, is designed to meet the requirements of the primary stage for a smaller receptive field size, thus maintaining efficient feature extraction. In feature extraction, the role of multi-rate depth dilated convolution is from complex to simple: it first extracts rich context information from complex feature maps, and then performs simple morphological filtering on the feature maps required for each concise expression to achieve dilation. In this way, simple regional feature maps promote the clarification and simplification of the learning process. The purpose of depth convolution is to make the learning process more organized and orderly. The overall network architecture diagram is as Figure 4 shown.

[0053] S4. Use the image data with defect annotations in step S2 to train the highway defect detection model to obtain a trained highway defect detection model;

[0054] S5. After using the inspection drone to take highway defect images, send the images into the trained highway defect detection model to complete defect detection.

[0055] In the embodiments of the present application, the system architecture design of the method according to the present application adopts a hierarchical method to ensure the decoupling and efficient communication between components. The system architecture mainly consists of the following layers:

[0056] (1) Data layer

[0057] MySQL: A relational database management system used to store structured data such as system configurations, user information, inspection tasks, and reports.

[0058] LFS (Large File Storage): Used to store large-scale image and video files collected by drones, optimizing the storage and retrieval of large amounts of data.

[0059] Redis: A high-performance key-value storage system used to cache frequently accessed data such as session information and temporary files to improve system performance.

[0060] (2) Service layer

[0061] Spring Cloud Bus: A message-based microservices communication framework used to implement configuration management and event delivery between services, ensuring high availability and consistency of services.

[0062] Microservices: The system is split into multiple independent microservices, each responsible for a specific function such as task management, data collection, AI analysis, etc.

[0063] (3) Communication layer

[0064] NGINX: A high-performance HTTP and reverse proxy server used for load balancing and reverse proxy to improve the throughput and response speed of the system.

[0065] Tomcat: As a Servlet container, used to run Java Web applications and handle HTTP requests and responses.

[0066] HTTP: As the main communication protocol, used for data exchange between the client and the server.

[0067] (4) Presentation layer

[0068] Web: A browser-based user interface that allows users to perform task planning, view reports, and manage configurations.

[0069] App: A mobile application that provides a convenient operation interface for on-site staff for real-time monitoring and on-site data collection.

[0070] System architecture working principle:

[0071] Data layer: Responsible for data storage and management. MySQL stores structured data such as system configurations, user information, and inspection tasks; LFS stores large-scale image and video files; Redis caches frequently accessed data to improve system performance.

[0072] Service layer: Composed of Spring Cloud Bus and microservices. Spring Cloud Bus is responsible for configuration management and event transmission between services, ensuring high availability and consistency of services; microservices are responsible for specific business functions such as task management, data collection, AI analysis, etc., realizing modular development.

[0073] Communication layer: Composed of NGINX, Tomcat and HTTP. NGINX performs load balancing and reverse proxy, improving system throughput and response speed; Tomcat acts as a Servlet container to run Java Web applications; HTTP serves as the main communication protocol to realize data exchange between the client and the server.

[0074] Presentation layer: Composed of Web and App. Web provides a browser-based user interface for managers to perform task planning, view inspection reports and manage configurations; App provides a mobile operation platform for on-site staff to conduct real-time monitoring and on-site data collection.

[0075] Improvements compared with the prior art:

[0076] Modular design: Splitting the system functions into multiple independent microservices improves the scalability and maintainability of the system.

[0077] Caching mechanism: Utilizing Redis to cache data reduces database access and improves system performance.

[0078] 5G network integration: Combining the high-speed data transmission ability of 5G network realizes real-time data processing and feedback, enhancing the response speed of inspections.

[0079] AI big data analysis: Using deep learning algorithms for defect identification improves the accuracy and efficiency of inspections.

[0080] 4. System integration and optimization

[0081] Automated inspection: Utilizing technologies such as AI and high-precision positioning to realize automated collection and monitoring of highway conditions.

[0082] Data analysis and early warning: Analyzing the collected data to promptly discover potential problems and issue early warnings.

[0083] Report generation and export: Automatically generating inspection reports according to format specifications, supporting one-key export for easy viewing and analysis by managers.

[0084] Real-time scheduling: According to inspection results, real-time scheduling of maintenance personnel is realized for quick response.

[0085] The present invention has the following advantages: High degree of automation: The system can achieve a fully automatic inspection process, from data collection to defect identification and then to report generation, significantly reducing manual intervention.

[0086] High recognition accuracy: By adopting a new deep learning model for object detection, the system shows high accuracy in defect identification, surpassing traditional methods.

[0087] Strong data processing ability: The system can process a large amount of inspection data and quickly generate inspection reports, improving the efficiency of data processing.

[0088] Strong environmental adaptability: The system can maintain high working performance under different weather and lighting conditions, with good robustness.

[0089] Economic benefits: By improving the inspection efficiency and accuracy, the system can reduce maintenance costs and minimize traffic accidents caused by improper road maintenance, thus saving economic losses.

[0090] Social benefits: The application of the system can enhance the safety of highways, reduce traffic accidents, and protect people's lives and property, having important social value.

[0091] Environmental benefits: By reducing the frequency of manual inspections, the system reduces energy consumption and environmental pollution, conforming to the concept of sustainable development.

[0092] In the embodiments of the present application, the actual defect detection process using the above method is as follows:

[0093] 1. Data collection and dataset division

[0094] The experiment uses a dataset constructed from photos of highway bridges and slopes taken by drones. This dataset collects 6 types of highway defect types, namely bridge biodegradation, bridge concrete leaching, bridge cracks, bridge efflorescence, exposed reinforcement steel of bridges, and slope landslides, with 500 photos for each type. Then, data augmentation is performed on each photo.

[0095] 2. Data annotation

[0096] The image data obtained through data augmentation needs to be defect-annotated for subsequent model training.

[0097] Select LabelImg to efficiently annotate the YOLO-format dataset. For each enhanced image, data annotation is required, and a bounding box needs to be drawn around it. The bounding box is usually represented by four numbers: (x_min, y_min, x_max, y_max), where (x_min, y_min) are the coordinates of the upper left corner of the bounding box, and (x_max, y_max) are the coordinates of the lower right corner. Assign a class label to each bounding box. This label should correspond to the previously defined classes. Before starting to train the model, ensure that all annotations are correct. Incorrect annotations may lead to a decline in model performance.

[0098] 3. Model Training and Validation

[0099] The labeled dataset will be divided into a training set of 2400 images, a validation set of 300 images, and a test set of 300 images in the ratio of 8:1:1.

[0100] The Ubuntu system is used as the operating system for this experiment. The GPU used is NVIDIA GeForce RTX4060 with 12GB of video memory, and the CPU is 12th Gen CoreTM i7-12700H. The deep learning framework is PyTorch version 1.13.1, the Python version is 3.8, and the CUDA version is 11.7. None of the network models use pre-trained models. During the training process, 300 training epochs are set, the batch_size is 16, and to balance the convergence speed and stability to a certain extent, the initial learning rate is set to 0.01. To reduce the computational overhead, the optimizer SGD is used, and the input image size is uniformly scaled to 640×640.

[0101] When evaluating the detection accuracy and speed of the model, common evaluation metrics include recall (R), precision (P), F1-score, average precision (AP), and mean average precision (mAP). Recall is used to represent the proportion of the number of correctly detected positive samples to the total number of positive samples in the detection task by the model:

[0102]

[0103] Precision is used to represent the proportion of the number of correctly detected samples to the total number of samples in the actual detection task by the model:

[0104]

[0105] Among them, TP represents the number of correct targets in the detection results, FP represents the number of incorrect targets in the detection results, and FN represents the number of missing targets among the correct targets.

[0106] The F1-score is usually used to represent the harmonic mean of precision and recall that needs to be measured simultaneously due to unreasonable data partitioning:

[0107]

[0108] The average precision (AP) is obtained from the area calculated according to the Precision-Recall curve:

[0109] AP = ∫0 1 P(R)dR (11)

[0110] The mean average precision (mAP) is the average of the average precisions of all classes:

[0111]

[0112] In addition, the evaluation metrics also include the model's computational complexity (GFLOPs), the number of model parameters (parameters), and the detection speed (frames per second, FPS). The model's computational complexity and the number of model parameters are key metrics for evaluating the model's computational efficiency and model capacity respectively; the detection speed can evaluate the ability to process the number of images per second and is an important criterion for measuring the model's performance.

[0113] After the comparative ablation experiment, the model can effectively improve the recognition accuracy and detection speed of highway defect targets. On the basis of YOLOv8s, the mAP, P, and R are improved by 16.1%, 9.3%, and 14.9% respectively. Example of test results:

[0114] 4. System Platform Design

[0115] The results obtained from training will be displayed on the platform designed in this system.

[0116] In this project, the design of the system platform is divided into two main parts: the WEB end and the APP end to meet the needs of different users in different scenarios.

[0117] WEB End Design

[0118] As the central hub of the system, the WEB end undertakes the core functions of task management, data processing, and report export, and at the same time integrates the data management modules of device management, personnel management, system management, and slope bridges.

[0119] Task Management: The WEB side allows administrators to create, assign, and monitor inspection tasks. The task management module provides functions such as task scheduling, priority setting, and status tracking.

[0120] Data Processing and Report Export: The WEB side is responsible for receiving data reported by the APP side and using AI big data technology to autonomously identify defects in highway inspections. This function is achieved by integrating advanced image recognition algorithms, which can perform in-depth analysis on images taken by drones and automatically identify defects in key parts such as bridges and slopes, such as cracks and damages. High-performance computing resources are deployed at the system backend, which can process large-scale image data and use deep learning models for fast and accurate defect identification. These models are trained with a large amount of labeled data to ensure high accuracy of the recognition results for further analysis and processing. The processed data can generate standardized inspection reports and support exports in multiple formats, such as PDF or Excel.

[0121] Equipment Management: Includes maintenance records, flight logs, and status monitoring of drones to ensure the optimal operation of the equipment.

[0122] Personnel Management: Manages the information of drone pilots, including qualification review, training records, and work assignment.

[0123] System Management: Responsible for system configuration, user permission management, and log auditing to ensure the security and stability of the system.

[0124] Slope and Bridge Data Management: Specifically manages the data of highway slopes and bridges, including historical data comparison, disease development tracking, and maintenance records.

[0125] APP Side Design

[0126] The APP side provides a mobile operation platform for on-site staff, making task receiving, task execution, data reporting, and flight safety monitoring more convenient.

[0127] Task Receiving: Pilots can receive tasks assigned by the WEB side through the APP side and view task details, including inspection areas, time requirements, and specific instructions.

[0128] Task Execution: The APP side provides guidelines for task execution, including flight path planning, shooting angles, and data collection standards.

[0129] Data Reporting: The data collected during task execution can be reported to the WEB side in real time through the APP side to ensure the timeliness of the data.

[0130] Flight Safety: The APP side integrates flight safety monitoring functions, including the real-time position of the drone, battery status, and emergency response measures.

[0131] 5. Model Deployment and Testing

[0132] Model Deployment

[0133] Model deployment is the process of integrating a trained machine learning model into a production environment so that it can process actual data and provide prediction services. The system first exports the trained model file (.pt file) and encapsulates the model into an API using the Flask framework for direct invocation by the server side.

[0134] Server Configuration:

[0135]

[0136] Test Experiment Design

[0137] A series of experiments are conducted to verify the effectiveness of the highway inspection and control system based on drones and AI big data. The experimental design follows a scientific and rigorous methodology to ensure the validity and reliability of the results. The experiments are divided into the following parts:

[0138] Experimental Purpose: To verify the performance of the system in actual highway inspections, including the integrity of data collection, the accuracy of AI defect identification, the real-time nature of data processing, and the convenience of user operation.

[0139] Experimental Environment: Representative highway sections are selected for on-site testing, including tests under different weather and lighting conditions, to evaluate the robustness of the system in various environments.

[0140] Experimental Method: Through comparative experiments, the data collected by drones is compared with the results of traditional manual inspections to evaluate the effectiveness of the system. At the same time, the accuracy of the AI defect identification function is tested, and the time for data processing and report generation is recorded.

[0141] Based on the comprehensive experimental results, the highway inspection and control system based on drones and AI big data shows good performance in aspects such as data collection, defect identification, data processing, and user operation. The implementation of the system has significantly improved the inspection efficiency, reduced the labor cost, and enhanced the timeliness and accuracy of road maintenance.

[0142] The above description illustrates and describes a preferred embodiment of the present invention. However, as previously mentioned, it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the above teachings or the techniques or knowledge in related fields. Any alterations and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A highway defect detection method for inspection drones, characterized by: The following steps are included S1. Use inspection drones to capture image data of multiple highway defect types and enhance the image data; S2. marking defects on the enhanced image data; S3. Build a highway defect detection model based on the improved YOLOv8n algorithm; S4. Train the highway defect detection model using the image data after defect annotation in step S2 to obtain a trained highway defect detection model; S5. After using the inspection drone to take images of highway defects, the images are sent to the trained highway defect detection model to complete defect detection.

2. The highway defect detection method for inspection drones according to claim 1 is characterized by: The multiple highway defect types include: bridge biodegradation, bridge concrete leaching, bridge cracks, bridge weathering, bridge steel bar exposure and slope landslides.

3. The highway defect detection method for inspection drones according to claim 1 is characterized by: In the step S1, multiple images need to be collected for each highway defect type.

4. The highway defect detection method for inspection drones according to claim 1 is characterized by: The manner of enhancing the image data includes: one or more of rotation, flipping, brightness enhancement, contrast enhancement and noise introduction.

5. The highway defect detection method for inspection drones according to claim 1 is characterized by: The step S2 comprises: Use LabelImg to annotate the YOLO format dataset: For each enhanced image to do data annotation, you need to draw a bounding box around it. The bounding box is represented by four numbers: (x_min, y_min, x_max, y_max), where (x_min, y_min) is the coordinate of the upper left corner of the bounding box, and (x_max, y_max) is the coordinate of the lower right corner; assign a category label to each bounding box, which corresponds to the defect category.

6. The highway defect detection method for inspection drones according to claim 1 is characterized by: In step S3, the improved YOLOv8n algorithm includes: (1) Based on the original YOLOv8n algorithm, the CIoU loss function is replaced by the EIoU loss function: The original YOLOv8n model uses CIoU as the bounding box regression function to calculate the similarity between bounding boxes in target recognition. In the improved YOLOv8n model, the CIoU loss function is improved to the EIoU loss function. (2) Based on the original YOLOv8n algorithm, PC2f is used to replace the two C2f modules in the front end of the original backbone: The original C2f module only uses ordinary Conv to extract image features, and PC2f is fused with C2f through a partial convolution PConv module; that is, Conv is only applied to a part of the input channel for spatial feature extraction, and the remaining channels remain unchanged; PC2f is used to replace the two C2f modules at the front end of the original backbone to accelerate the training and reasoning process of the model and improve the efficiency of the model; (3) Introducing the attention mechanism into the backbone: The DWRSeg module combined with the attention mechanism is introduced, and the two C2f modules at the end of the original network backbone are replaced by the DWRSeg module. The convolution layer is retained after each DWRSeg, and the final output is introduced into the SPPF module for processing. The SPPF module uses pools with different kernel sizes to combine feature maps, and then passes the result to the neck layer.

Images