基于ResNet-SSD的铁路扣件缺损检测方法  

A method of railway fastener defect detection based on ResNet-SSD

作　　者：王登飞 苏宏升[1] 陈登科 赵小娟 WANG Dengfei;SU Hongsheng;CHEN Dengke;ZHAO Xiaojuan(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Key Laboratory of Plateau Traffic Information Engineering and Control of Gansu Province,Lanzhou 730070,China;BaotouVehicle Depot,China Railway Hohhot Group Co.,Ltd.,Hohhot 010000,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070 [2]甘肃省高原交通信息工程及控制重点实验室,甘肃兰州730070 [3]中国铁路呼和浩特局集团公司包头车辆段,内蒙古呼和浩特010000

出　　处：《Journal of Measurement Science and Instrumentation》2023年第3期360-368,共9页测试科学与仪器（英文版）

基　　金：National Natural Science Foundation of China(Nos.61867003,62141303);Science and Technology Project of Gansu Province(No.21ZD4WA018)。

摘　　要：铁路轨面扣件的缺失、损坏状态与行车安全密切关系。为进一步提高铁路扣件检测性能,提出基于残差网络-单发多框检测器(Residual network-single shot multibox detector,ResNet-SSD)的铁路扣件缺损检测方法。首先,采用虚拟图像技术进行扣件缺损状态图像建模来扩展扣件样本特征信息,完善样本库以解决各类扣件图像不均衡问题。其次,使用ResNet构建特征提取网络,SSD模型进行目标检测,完成对扣件缺损状态的检测。实验结果表明,采用虚拟图像扩展现场采集扣件图像集后,ResNet-SSD网络有更快的收敛速度且mAP提高了3.06%(95.47%-98.53%);ResNet-SSD在召回率为98.8%时,检测准确度可达98.5%,达到了在交并比(Intersection over union,IoU)为0.5时较VGG16-SSD高、与Faster-RCNN相同的评价指标;就检测速度而言,该方法为Faster-RCNN的6倍,VGG16-SSD的1.4倍。该方法可满足铁路扣件实时自动检测的需求。Missing and broken states of railway fasteners are directly related to driving safety.A deep network image detection method for railway fasteners based on residual network-single shot multibox detector(ResNet-SSD)is proposed to improve detection performance.Firstly,virtual image technology is used to simulate the defect states of fasteners,so as to expand feature information and overcome the imbalance problem of various fastener images.Secondly,ResNet is used to construct feature extraction network,and SSD is used for target detection to form ResNet-SSD network to detect the defect states of fasteners.Experimental results show that the ResNet-SSD has faster convergence speed and mean average precision(mAP)is increased by 3.06%from 95.47%to 98.53%after using the virtual image to expand the field collected fastener set;The detection accuracy of ResNet-SSD can reach 98.5%when the recall is 98.8%,which is higher than that of VGG16-SSD and the same as that of Faster-RCNN when intersection over union(IoU)is 0.5;The detection speed of this method is 6 times that of Faster RCNN and 1.5 times that of VGG16-SSD.Therefore,this method can meet the needs of real-time detection of railway fasteners.

关 键 词：铁路扣件 虚拟图像 图像识别 残差网络 单发多框检测器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U216.3[自动化与计算机技术—计算机科学与技术]