基于多层级特征融合的钢轨表面伤损检测方法  

Damage Detection Method for Rail Surface Based on Multi-Level Feature Fusion

作　　者：韩强 刘俊博 冯其波[1] 王胜春 戴鹏 HAN Qiang;LIU Junbo;FENG Qibo;WANG Shengchun;DAI Peng(School of Science,Beijing Jiaotong University,Beijing 100044,China;Infrastructure Inspection Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China)

机构地区：[1]北京交通大学理学院,北京100081 [2]中国铁道科学研究院集团有限公司基础设施检测研究所,北京100081

出　　处：《中国铁道科学》2021年第5期41-49,共9页China Railway Science

基　　金：中国铁道科学研究院集团有限公司院基金课题(2019YJ158)。

摘　　要：现有的钢轨表面伤损检测方法存在鲁棒性差、误检率高和容易漏检小面积伤损区域的问题。为此,提出一种基于多层级特征融合的钢轨表面伤损检测方法。首先,利用高速综合检测车搭载轨道图像采集系统,在实际的铁路线路采集轨道图像,并对表面伤损进行人工标注;然后,在钢轨图像数量有限的情况下,利用钢轨表面伤损数据集构建策略,提升训练样本图像的数量和多样性;最后,利用上述数据集,训练基于多层级特征融合的目标检测网络,实现钢轨表面伤损区域的自动检测。将所提新方法与现有方法进行对比试验,结果表明:新方法在钢轨表面伤损数据集上具有最优性能,实现了端到端的钢轨表面伤损检测,能够满足实际应用需求。The existing detection methods for rail surface damage have the problems of poor robustness,high false detection rate and the misdetection of small damage areas.Therefore,a damage detection method for rail surface based on multi-level feature fusion is proposed.Firstly,the track image acquisition system of high-speed comprehensive inspection vehicle is used to collect the track images on the actual railway line,and the surface damages are manually marked.Secondly,a dataset construction strategy for rail surface damage is used to improve the number and diversity of training sample images with the limited number of rail images.Finally,the above datasets are used to train the target detection network based on multi-level feature fusion to realize the automatic detection for rail surface damage areas.The comparison tests are performed between the new proposed method and the existing methods.The results show that the new method has the optimal performance on the rail surface damage datasets,realizing the end-to-end rail surface damage detection,which can meet the practical application requirements.

关 键 词：钢轨表面伤损 多层级特征融合 小目标检测 深度卷积神经网络 

分 类 号：U216.3[交通运输工程—道路与铁道工程]