基于YOLOv3的金属表面缺陷检测研究  

Research on Metal Surface Defect Detection Based on YOLOv3

作　　者：任伟建[1,2] 陈明文 康朝海[1,2] 霍凤财[1,2] 任璐 张永丰 REN Weijian;CHEN Mingwen;KANG Chaohai;HUO Fengcai;REN Lu;ZHANG Yongfeng(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China;Heilongjiang Provincial Key Laboratory of Networking and Intelligent Control,Northeast Petroleum University,Daqing 163318,China;Offshore Oil Engineering Co.,Ltd.,Tianjin 300450,China;Institute of Planning and Design of NO.2 Oil Production Plant,Daqing Oilfield Co.,Ltd.,Daqing 163318,China)

机构地区：[1]东北石油大学电气信息工程学院,黑龙江大庆163318 [2]东北石油大学黑龙江省网络化与智能控制重点实验室,黑龙江大庆163318 [3]海洋石油工程股份有限公司,天津300450 [4]大庆油田有限责任公司第二采油厂规划设计研究所,黑龙江大庆163318

出　　处：《控制工程》2024年第7期1219-1228,共10页Control Engineering of China

基　　金：国家自然科学基金资助项目(61933007,61873058)。

摘　　要：为了解决金属表面缺陷检测的漏检、误检等问题,提出了一种改进YOLOv3算法。首先,使用动态激活函数替换主干特征提取网络中所有残差块的激活函数,并加入了混合注意力机制,强化其对复杂缺陷目标的特征提取能力。然后,在特征金字塔网络部分新增一个104×104的特征层,并将浅层网络与深层网络进行逐层特征融合,增强算法对小缺陷目标检测的敏感性。最后,利用K-Means++聚类算法替换K-Means聚类算法,筛选出适用于金属表面缺陷检测的最优先验框尺寸,使目标定位更加准确。实验结果表明,改进YOLOv3算法的每秒检测帧数(frames per second,FPS)可达到32.3,平均精度均值(mean average precision,mAP)可达到78.69%,检测性能得到了明显提升。In order to solve the problems of missing and false detection in metal surface defect detection,an improved YOLOv3 algorithm is proposed.Firstly,a dynamic activation function is used to replace the activation functions of all residual blocks in the backbone feature extraction network,and a hybrid attention mechanism is added to strengthen its feature extraction ability for complex defect targets.Then,a 104×104 feature layer is added to the feature pyramid network part,and the shallow network features and deep network features are fused layer by layer to enhance the sensitivity of the algorithm to small defect target detection.Finally,the K-Means++clustering algorithm is used to replace the K-Means clustering algorithm to screen out the most priority priori box size suitable for metal surface defect detection,so that the target positioning is more accurate.Experimental results show that the frames per second(FPS)of the improved YOLOv3 algorithm can reach 32.3,the mean average precision(mAP)can reach 78.69%,and the detection performance is significantly improved.

关 键 词：缺陷检测 特征提取网络 损失函数 特征金字塔网络 先验框尺寸 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]