Research on deep learning method for rail surface defect detection

Rail surface defect detection plays a critical role in the maintenance of the rail transportation system. Video analysis technology is a promising method to detect defects due to its low cost and effectiveness. Several attempts with hand-craft features have been made to obtain the detection results by using traditional machine vision algorithms. However, these methods suffer from imprecise results due to challenging conditions, such as deteriorated and changeable lighting environment and various types of complex rail surface defects. Recently, classification methods with complex deep convolutional networks have become popular. Despite their high accuracy, these methods cannot meet the requirements of defects localisation and real-time processing in practice. To solve these problems, this study proposes a novel object detection algorithm to detect rail defects. The net architecture of the proposed algorithm includes a backbone network using MobileNet and several novel detection layers with multi-scale feature maps inspired by you only look once (YOLO) and feature pyramid networks. Two different architectures of MobileNet are used to estimate the performance of defects detection. The experimental results demonstrate the great potential of the proposed algorithm with fast inference speed and high accuracy in the industry.