基于改进YOLOv5的轴承表面缺陷检测  

Detection of Bearing Surface Defects Based on Improved YOLOv5

作　　者：吴迪[1] 于正林[1,2] 徐式达 周斌 邵长顺 WU Di;YU Zhenglin;XU Shida;ZHOU Bin;SHAO Changshun(School of Mechanical and Electrical Engineering,Changchun University of Science and Technology,Changchun 130022,China;Chongqing Research Institute,Changchun University of Science and Technology,Chongqing 401135,China)

机构地区：[1]长春理工大学机电工程学院,长春130022 [2]长春理工大学重庆研究院,重庆401135

出　　处：《组合机床与自动化加工技术》2024年第6期166-170,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：吉林省科技发展计划基金项目(20190302069GX);吉林省科技厅基础研究项目(202002044JC)。

摘　　要：传统的轴承表面缺陷检测由于缺陷目标较小,错检漏检率高,检测效率低等问题,为此提出一种基于YOLOv5网络改进的算法模型。首先,在主干网络中添加高效通道注意力机制(efficient channel attention, ECA),增强网络的特征提取能力,集中关注各种影响轴承质量的重点信息;其次,在YOLOv5网络基础上添加小目标检测层,通过补充融合特征层和引入额外检测头,提高网络对小目标缺陷检测的精度;最后,在特征融合网络中,融入简化后的加权双向特征金字塔网络(bidirectional feature pyramid network, BiFPN),在不增加较多计算成本的基础上,更好地实现多尺度特征融合。在构建的深沟球轴承表面缺陷数据集上的实验结果表明,相比于原YOLOv5s模型,精确率、召回率、平均精度分别提高了5.8%、2.4%、5.3%,检测速度为71 f/s,满足工业大批量检测的要求。Traditional bearing surface defect detection has problems such as small defect targets,high false or missed detection rates,and low detection efficiency,therefore,an improved algorithm model based on YOLOv5 network is proposed.Firstly,add an efficient channel attention(ECA)mechanism to the backbone network to enhance the network′s feature extraction ability and focus on various key information that affects bearing quality;Secondly,a small object detection layer is added to the YOLOv5 network,and the accuracy of small object defect detection is improved by supplementing the fusion feature layer and introducing additional detection heads;Finally,in the feature fusion network,a simplified bidirectional feature pyramid network(BiFPN)is incorporated to better achieve multi-scale feature fusion without increasing computational costs.The experimental results on the constructed deep groove ball bearing surface defect dataset show that compared to the original YOLOv5s model,the accuracy,recall,and average accuracy have been improved by 5.8%,2.4%,and 5.3%,respectively,with a detection speed of 71 f/s,meeting the requirements of industrial mass inspection.

关 键 词：YOLOv5 缺陷检测 注意力机制 小目标检测层 简化BiFPN 

分 类 号：TH16[机械工程—机械制造及自动化] TG66[金属学及工艺—金属切削加工及机床]