摘要
针对复杂背景下铁路扣件外观缺陷形态检测困难的问题,提出一种改进的YOLOv5s(RFYOLOv5s)铁路扣件检测算法。首先,结合铁路扣件形态特征,采用DenseBlock模块代替YOLOv5s网络中的Focus模块,并嵌入Ghost-BottleNeck模块,替换YOLOv5s主干网络中的Bottleneck模块;然后,引入损失函数LCIoU对YOLOv5s网络进行优化。此外,对原网络中马赛克数据增强方式进行改进,以丰富数据样本特征。将RF-YOLOv5s算法与Faster R-CNN,SSD,YOLOv3,YOLOv4和YOLOv5s等经典目标检测算法对无人机获取的铁路扣件外观缺陷形态数据影像进行检测效果对比。结果表明:RF-YOLOv5s算法的平均检测精确率可达95.46%,平均检测用时为15 ms,可实现对昏暗、遮挡、杂物干扰和模糊复杂背景下铁路扣件外观缺陷形态的准确检测;相比于几种经典目标检测算法,RF-YOLOv5s算法的整体性能表现更为出色,能够满足在实际铁路工况下对扣件实时检测的需求,为铁路管养部门提供技术参考。
In order to solve the problem of the difficulty in detecting appearance defect morphology of railway fasteners under complex background,an improved YOLOv5s(RF-YOLOv5s)railway fastener detection algorithm is proposed.First,combined with the morphological characteristics of railway fasteners,DenseBlock module is used to replace the Focus module in the YOLOv5s network,and the Ghost-BottleNeck module is embedded to replace the Bottleneck module in the YOLOv5s backbone network.Then,loss function LCIoU is introduced to optimize the YOLOv5s network.In addition,the enhancement method of mosaic data in the original network is improved to enrich the characteristics of data samples.The proposed RF-YOLOv5s algorithm is compared with the classical object detection algorithms,including Faster R-CNN,SSD,YOLOv3,YOLOv4 and YOLOv5s,to carry out the comparison of detection results of the morphological data images of railway fastener appearance defects obtained by UAV.The results show that the average detection accuracy of RF-YOLOv5s algorithm can reach 95.46%and the average detection time is 15 ms.The proposed algorithm can achieve accurate detection of the appearance defect morphology of railway fasteners under fuzzy and complex backgrounds,including low light,barrier and block,and sundry interference.Compared with several classical object detection algorithms,the overall performance of RF-YOLOv5s algorithm has a better performance,which can meet the requirements of real-time detection of fasteners under practical working conditions on railways,and provide technical reference for railway management and maintenance departments.
作者
吴送英
刘林芽
江家明
张洪
左志远
WU Songying;LIU Linya;JIANG Jiaming;ZHANG Hong;ZUO Zhiyuan(School of Transportation Engineering,East China Jiaotong University,Nanchang Jiangxi 330013,China)
出处
《中国铁道科学》
EI
CAS
CSCD
北大核心
2023年第3期53-63,共11页
China Railway Science
基金
国家自然科学基金资助项目(51968025)
江西省交通运输厅科技重点项目(2022Z0003)
江西省教育厅科学技术研究重点项目(GJJ210603,GJJ171287,GJJ204613)
江西省研究生创新基金(YC 2021-B143)。