Rail Surface Defect Detection Method Based on YOLOv3 Deep Learning Networks

Abstract

In order to ensure the rapid, safe and stable operation of trains, it is very important to detect the flaws on the rail surface. At present, although there are many detection methods for rail surface defects, the comprehensiveness, rapidity and accuracy of defect detection are still not satisfactory. Therefore, this paper presents a deep learning method using the YOLOv3 algorithm to realize rail surface defect detection. It first resets the input rail image size to 416*416, and then divides the rail image into S*S cells. According to the position of the defect in the cell, the width and height of the defect and the coordinates of the center point are calculated by the method of dimensional clustering, the coordinates are normalized. At the same time, it uses logistic regression to predict the bounding box object score, the binary cross-entropy loss is used to predict the categories that the bounding box may contain, the confidence is calculated and then prediction. The test results show that the recognition rate of this algorithm can reach more than 97%, and the identification time is about 0.15s. This method has great advantages for the detection of rail surface defects.