基于改进YOLOv7的钢轨螺孔伤损检测方法  

Detection Method of Rail Screw Hole Damage Based on Improved YOLOv7

作　　者：许贵阳 付伟 张玉华 白堂博 XU Guiyang;FU Wei;ZHANG Yuhua;BAI Tangbo(School of Mechanical-Electronic and Vehicle Engineering,Beijing University of Civil Engineering and Architecture,Beijing100044,China;Beijing Key Laboratory of Performance Guarantee on Urban Rail Transit Vehicles,Beijing University of Civil Engineering and Architecture,Beijing100044,China;Infrastructure Inspection Research Institute,CARS,Beijing 100081,China)

机构地区：[1]北京建筑大学机电与车辆工程学院,北京100044 [2]北京建筑大学城市轨道交通车辆服役性能保障北京市重点实验室,北京100044 [3]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：《铁道建筑》2024年第3期52-58,共7页Railway Engineering

基　　金：国家自然科学基金(51975038);北京市自然科学基金(L211007,L221027)。

摘　　要：针对目前钢轨螺孔伤损识别过程中算法的检测精度较低、检测时间长、参数量大、计算资源占用高等问题,提出一种基于YOLOv7网络的改进算法模型。通过建立B显图像数据集,统计分析螺孔B显图像中待检测目标的面积大小、宽高比和占比情况,对YOLOv7网络进行改进,包括改进目标检测层、引入轻量级卷积、增加坐标注意力机制、优化损失函数,降低了改进算法的参数量和计算量,提高了网络的识别能力和检测速度。为评估该改进方法的有效性,进行了消融试验,并与Faster R-CNN(Regionbased Convolutional Neural Networks)、YOLOv3及YOLOv5算法进行了对比测试。结果表明,提出的改进YOLOv7算法综合表现优于其他算法,具有更高均值平均精度和更小的错检率,最终算法的均值平均精度为97%,参数量为18.0×10^(6),计算量为58.1×10^(9),检测时间为7.4 ms,能够较好地应用于钢轨螺孔伤损检测场景。A modified algorithm model based on YOLOv7 network was proposed to address the issues of low detection accuracy,long detection time,large parameter count,and high computational resource consumption in the current process of identifying steel rail screw hole damage.By establishing a B-display image dataset and statistically analyzing the area size,aspect ratio,and proportion of the target to be detected in the screw hole B-display image,the YOLOv7 network was improved,including improving the target detection layer,introducing lightweight convolution,increasing coordinate attention mechanism,optimizing loss function,reducing the parameter and computational complexity of the improved algorithm,and improving the network's recognition ability and detection speed.To evaluate the effectiveness of the improved method,ablation experiments were conducted and compared with Faster R-CNN(Region based Convolutional Neural Networks),YOLOv3,and YOLOv5 algorithms.The results show that the proposed improved YOLOv7 algorithm performs better than other algorithms in terms of overall performance,with higher mean average accuracy and smaller false detection rate.The final algorithm has a mean average accuracy of 97%,a parameter size of 18.0×10^(6),a computational complexity of 58.1×10^(9),and a detection time of 7.4 ms.It can be well applied to the detection of steel rail screw hole damage scenarios.

关 键 词：高速铁路 钢轨探伤 目标检测 螺孔伤损 YOLOv7 B显图像 

分 类 号：U216.3[交通运输工程—道路与铁道工程] TH18[机械工程—机械制造及自动化]