基于YOLOv5和ConvNext的钢铁表面缺陷检测研究  

Research on Steel Surface Defect Detection Based on YOLOv5 and ConvNext

作　　者：李强强 皋军[2] 邵星[2] 王翠香[2] LI Qiangqiang;GAO Jun;SHAO Xing;WANG Cuixiang(School of Mechanical Engineering,Yancheng Institute of Technology,Yancheng 224000,China;School of Information Engineering,Yancheng Institute of Technology,Yancheng 224000,China)

机构地区：[1]盐城工学院机械工程学院,盐城224000 [2]盐城工学院信息工程学院,盐城224000

出　　处：《组合机床与自动化加工技术》2024年第10期160-165,170,共7页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(62076215,61502411);教育部新一代信息技术创新项目(2020ITA02057);盐城工学院研究生科研与实践创新计划项目(SJCX22_XZ035,SJCX22_XY061)。

摘　　要：为解决工业钢铁表面缺陷检测速度慢、准确度低问题,提出一种基于改进YOLOv5网络的检测方法。在YOLOv5网络的FPN特征金字塔模块中加入ECANet模块,以提高检测精度;利用K-Means算法在NEU-DET数据集上重新聚类,生成3组新的先验框,降低网络损失;针对钢铁缺陷的小目标特征,将ConvNext网络应用到YOLOv5的主干网络中,用ConvNext网络提取小目标缺陷特征,增强模型学习能力。实验结果表明,改进后的YOLOv5模型与原YOLOv5模型相比,mAP提升了3.84%,平均检测速率为36.9 frame/s,能够做到快速和准确的检测,满足实际应用需求。In order to solve the problem of slow speed and low accuracy of surface defect detection of industrial steel,a detection method based on improved YOLOv5 network was proposed.The ECANet module is added to the FPN feature pyramid module of YOLOv5 network to improve detection precision;The K-Means algorithm is used to recluster the NEU-DET data set,generate three new sets of prior boxes,and reduce the network loss;Aiming at the small target features of steel defects,ConvNext network is applied to the backbone network of YOLOv5,and ConvNext network is used to extract the small target defect features and enhance the model learning ability.The experimental results show that compared with the original YOLOv5 model,the map of the improved YOLOv5 model is increased by 3.84%,and the average detection rate is 36.9 frame/s,which can achieve fast and accurate detection and meet the practical application requirements.

关 键 词：缺陷检测 K-MEANS算法 ConvNext ECANet YOLO 

分 类 号：TH162[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]