基于深度学习的轨道表面缺陷检测  

Rail surface defect detection based on deep learning

作　　者：胡璐萍 管声启[2] 王剑楠 吴哲 刘懂懂 HU Luping;GUAN Shengqi;WANG Jiannan;WU zhe;LIU Dongdong(School of Mechanical and Electrical Engineering,Xi'an Traffic Engineering Institute,Xi'an 710300,China;School of Mechanical and Electrical Engineering,Xi'an Polytechnic University,Xi'an 710048,China;Xi'an Aeronautical Polytechnic Institute,Xi'an 710089,China)

机构地区：[1]西安交通工程学院机械与电气工程学院,陕西西安710300 [2]西安工程大学机电工程学院,陕西西安710048 [3]西安航空职业技术学院,陕西西安710089

出　　处：《甘肃科学学报》2025年第1期16-24,共9页Journal of Gansu Sciences

基　　金：陕西省教育厅科研计划项目(23JK0531)。

摘　　要：针对钢轨表面缺陷检测精度不足和速度缓慢的难题,提出一种基于改进的YOLOv5s钢轨表面缺陷检测算法。在此算法中,首先在YOLOv5s网络中Bacbone部分内嵌入CBAM注意力模块,它能够有效挖掘钢轨损伤的通道和空间特征信息;同时,采用轻量化的CARAFE模块取代原YOLOv5s网络中Neck部分的传统上采样模块,避免上采样过程中特征信息的丢失,生成更多的细节和平滑的边缘,有效增加模型的感受域,这些优化措施显著提升了模型对钢轨缺陷特征的捕捉力,进而增强了模型的检测精度;其次,通过将YOLOv5s中的CBL模块更换为更加高效的GSConv卷积模块,实现了计算成本的节约和检测速度的提升。实验结果表明:改进后的YOLOv5s对钢轨表面缺陷检测平均精度mAP为90.98%,相比于YOLOv5s提升了2.7%,其中检测摩擦缺陷的平均精度提升了5%,对于疤痕缺陷和裂纹缺陷精度提高了2.3%和0.9%,检测速度提高了7.52帧/s,能够有效解决钢轨表面缺陷检测准确率低及检测速度慢的问题,证明了钢轨表面缺陷检测算法的有效性。Aiming at the problem of insufficient accuracy and slow speed of rail surface defect detection,a rail surface defect detection algorithm based on improved YOLOv5s was proposed.In this algorithm,firstly,CBAM attention module is embedded in Bacbone part of YOLOv5s network,which can effectively mine the channel and spatial characteristic information of rail damage.At the same time,the lightweight CARAFE module is adopted to replace the traditional sampling module in the Neck part of the original YOLOv5s network,so as to avoid the loss of feature information during the upsampling process,generate more details and smooth edges,and effectively increase the model receptive domain.These optimization measures significantly improve the model ability to capture rail defect features,and thus enhance the detection accuracy of the model.Secondly,by replacing the CBL module in YOLOv5s with the more efficient GSConv convolution module,the calculation cost is saved and the detection speed is improved.Finally,the experimental results show that the average accuracy mAP of the improved YOLOv5s for rail surface defects detection is 90.98%,which is 2.7%higher than that of YOLOv5s.The average accuracy of friction defects detection is increased by 5%,the accuracy of scar defects and crack defects are increased by 2.3%and 0.9%,and the detection speed is increased by 7.52 frames per second.It can effectively solve the problems of low accuracy and slow detection speed of rail surface defect detection,which shows the effectiveness of rail surface defect detection algorithm.

关 键 词：钢轨表面缺陷 YOLOv5s GSconv模块 轻量级采样模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]