A Real-time System of Two-stage Track Component Classification based on YOLOX-nano and ResNet34

Abstract

Due to the complex track environment, the use of only one stage for track component identification is prone to inaccurate component positioning, resulting in misjudgment of the target coordinate system and other problems. This paper proposes a two-stage recognition method based on YOLOX-nano and ResNet34, hoping to solve the problem of inaccurate component positioning in the existing classification system and also improve the recognition accuracy. In the first stage, the entire image is preliminarily screened through YOLOX-nano, so that the system can understand the image structure, obtain the possible range of components, and then obtain the leftmost and rightmost positions of the track through Hough Transform. Next, calculate the intersection with the sleeper range obtained in the first stage, and calculate the possible relative position of the component base on the intersection, thereby locking the range where the component is located, and handing this range to ResNet34 in the second stage for component defect detection.