Research on Real-Time Detection System of Rail Surface Defects Based on Deep Learning
Wang Yaodong, Hang Yu, Guo Baoqing, Hongmei Shi, Zujun Yu
Abstract
The heavy workload of rail track inspection makes it time consuming, and thus calls out a real-time inspection algorithm to achieve precise and efficient detiction. In this study, we developed a real-time detection system for rail surface. Our system utilizes machine vision and real-time algorithms to ensure efficient and fast inspections. Edge computing device is used for real-time detection of track defect. To increase detection accuracy and speed, we optimized the YOLOv5 structure by introducing depth-separable convolution and re-parameterization methods. Through training and evaluating the model on a dataset of rail surface defects, we achieved a mean average precision (mAP) of 83.2% and a detection speed of 51 FPS on edge computing devices. The performance of model outstrips that of other one-stage algorithms and backbone network detection results, as it exhibits high accuracy and speed. This achievement lays the groundwork for realizing real-time detection of rail defects and augmenting railroad safety.