融合Channel-Attention机制的金属表面缺陷检测算法  

Metal Surface Defect Detection Algorithm Fused with Channel-Attention Mechanism

作　　者：符秦沈 王桂棠[1,2] 张巧芬 邓宇平[1] FU Qin-shen;WANG Gui-tang;ZHANG Qiao-fen;DENG Yu-ping(School of Electromechanical Engineering,Guangdong University of Technology,Guangdong Guangzhou 510006,China;Foshan Cangke Intelligent Technology Co.,Ltd.,Guangdong Foshan 528225,China)

机构地区：[1]广东工业大学机电工程学院,广东广州510006 [2]佛山沧科智能科技有限公司,广东佛山528225

出　　处：《机械设计与制造》2023年第6期167-171,共5页Machinery Design & Manufacture

基　　金：国家自然科学基金(61705045)。

摘　　要：传统的缺陷检测方法存在各种各样的弊端,近年用基于深度学习缺陷检测方法成为研究热点。针对目前主流的目标检测算法需要牺牲速度以获取精度的问题,提出了一种融合Channel-Attention机制的SSD目标检测算法。该算法利用通道注意力机制来学习特征通道之间的关系,从而对特征层的每一个通道特征进行权重的分配,进而提升网络的学习能力。在铝型材外观数据集上的实验结果表明,该算法的检测性能达到了较好的效果。改进模型的平均精确均值达到了78.17%,与基础模型SSD相比,提升了3.55%。同时检测速度达到了45frame/s,在提升精度的同时保证了检测速度。验证了融合Channel-Attention机制在提升模型精度的同时没有给模型增加过多的计算量,满足工业实时性检测的要求。Traditional defect detection methods have various disadvantages.In recent years,the use of deep learning-based defect detection methods has become a research hotspot.Aiming at the problem that the current mainstream target detection algorithms need to sacrifice speed to obtain accuracy,an SSD target detection algorithm fused with Channel-Attention mechanism is proposed.The algorithm uses the channel attention mechanism to learn the relationship between feature channels,so as to assign weights to each channel feature in the feature layer,thereby improving the learning ability of the network.The experimental results on the aluminum profile appearance data set show that the detection performance of the algorithm has achieved good results.The mean average precision of the improved model reached 78.17%,which is an increase of 3.55%compared with the basic model SSD.At the same time,the detection speed has reached 45 frame/s,which ensures the detection speed while improving the accuracy.It is verified that the fusion Channel-Attention mechanism does not add too much calculation to the model while improving the accuracy of the model,and meets the requirements of industrial real-time detection.

关 键 词：缺陷检测 深度学习 注意力机制 多尺度检测 

分 类 号：TH16[机械工程—机械制造及自动化] TP391[自动化与计算机技术—计算机应用技术]