Abstract:

Rail surface defect detection is the important part of railway daily detection. According to the requirements of modern railway automatic detection technology for real-time detection and adaptability, a complete system of rail surface defect identification and analysis is constructed. According to the basic principle of machine vision, an image acquisition device with LED auxiliary light source and light shielding box is designed, and the collected image is manually labeled, and a relatively large rail surface defect data set with semantic segmentation annotation is established. The advanced semantic segmentation technology is applied to rail image analysis. Using a cascaded self coding structure（CASAE） semantic segmentation network, the defect image is transformed into a pixel level prediction mask based on semantic segmentation, and the segmentation results are classified by the compact Convolutional Neural Network（CNN）, so as to realize the recognition and classification of rail surface defects. An intelligent human-computer interaction system is built and tested by simulation experiment. The experimental results show that the detection accuracy of the system is more than 90%, and the average processing time of each image is 245.61 ms, which can replace the manual detection to some extent, and realize the digital management of rail defects.

Key words: real time detection, machine vision, semantic segmentation, rail defect detection system

摘要：

钢轨表面缺陷检测是铁路日常检测的重要部分，根据现代铁路自动化检测技术对实时检测和适应性的要求，构建了一个完整的钢轨表面缺陷识别和分析系统。根据机器视觉的基本原理，设计了一种带有LED辅助光源和遮光箱的图像采集装置，并将采集到的图像进行人工标注，建立了一个较为庞大的具有语义分割标注的钢轨表面缺陷数据集；将高级语义分割技术应用于钢轨图像分析，利用一种级联自编码结构（CASAE）的语义分割网络，将缺陷图像转化为基于语义分割的像素级预测掩码，并通过紧凑型卷积神经网络（CNN）将分割结果进行分类，从而实现钢轨表面缺陷的识别与分类；构建了智能化的人机交互系统，并将系统通过仿真实验的方式进行测试。实验结果表明，系统的检测准确率达到90%以上，每幅图像的平均处理时间为245.61 ms，可以在一定程度上代替人工检测，实现对钢轨缺陷的数字化管理。

关键词: 实时检测, 机器视觉, 语义分割, 钢轨缺陷检测系统