Two-Stage Key Component Recognition Algorithm for High-Speed Trains Based on YOLOv8

This paper addresses the recognition problem of key components (12 types of components including axle box covers, air springs, etc.) in trains. By connecting the output layer of YOLOv8's large component recognition and the input layer of YOLOv8's small component recognition, a YOLO-YOLO algorithm framework is proposed. After observation, it is found that the pixels of small components only account for an average of 0.25% of the original image pixels. Hence, this paper proposes to first recognize large components, and then use an automatic cropping algorithm to input the cropped large components into the next layer of network structure to recognize the small components therein, forming a two-stage algorithm. By comparing the recognition accuracy of Faster R-CNN and YOLOv8 for large images and the recognition rate for small images, it is found that YOLOv8's recognition accuracy for all components is higher than that of Faster R-CNN, and its results are more precise than those of the one-stage algorithm. The detection effect is better than the original algorithm, and the detection rate is higher than 95%.