Defect Detection of Rail Surface with Deep Convolutional Neural Networks

Abstract

With the heavy railway transportation pressure, the rail surface defect that came into being is an unavoidable problem, which is related to the railway transport safety. Therefore various defect detection methods of rail surface are proposed, but the accuracy, rapidity, stability and intelligence are still unsatisfactory. To overcome these difficulties, the paper proposes a deep convolution neural networks of the SegNet architecture to detect the surface defects of rail. Rail surface defect images are obtained by the system and sent to a 59 layers training networks designed by 120 rail training images to detect the rail surface defects. Compared with the traditional image threshold segmentation methods, this training networks achieve high efficiency, high accuracy and non-interference detection.