Research on rail surface defect detection algorithm based on improved yolov8

In this paper, a rail surface defect detection and classification algorithm based on the improved YOLOv8 is proposed, aiming to improve the accuracy and efficiency of defect detection in the rail maintenance process. First, on the basis of the theoretical research related to YOLOv8 algorithm, a rail surface defect classification algorithm is proposed, which utilizes model training to complete the task of accurate positioning and classification of rail surface defects to ensure the reliability of detection under complex working conditions. Secondly, the edge detection technology is utilized for straight line detection, and the results of straight line detection are used to extract the rail surface area in the panoramic image, and the experimental dataset under part of the real environment is constructed. The high quality of this dataset ensures the accuracy of the subsequent detection task and provides more realistic detection data support. Finally, in order to enhance the network’s ability to sense the global information of defects, the YOLOv8 network structure is improved. A multi-head self-attention layer is embedded at the end of its backbone network, and a cross-layer weighted cascade method is used to connect the shallow and deep networks, which effectively avoids information loss in the feature extraction process and improves the detection efficiency at the same time. The experimental results show that on the dataset of this paper, the improved algorithm improves mAP50 by $14.8 \%$ and $\mathrm{mAP} 50-95$ by $21.4 \%$ compared with the original algorithm. The detection speed of the algorithm reaches 84 FPS, which is able to accurately detect concrete surface crack targets in real time, and the effectiveness and superiority of the algorithm is verified by comparing it with other mainstream target algorithms.