融入协调注意力的MobileNetV2模型对铝带材表面缺陷识别的研究  

Research on MobileNetV2 aluminum strip surface defect recognition algorithm integrated with coordinated attention

作　　者：徐豫 李勇[1,2] 李贝贝 余凤智 任安吉 李家栋 XU Yu;LI Yong;LI Beibei;YU Fengzhi;REN Anji;LI Jiadong(State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,China;Guangxi Advanced Aluminum Processing Innovation Center Co.,Ltd.,Nanning 530007,China;Guangxi Alnan Aluminium Inc.,Nanning 530000,China)

机构地区：[1]东北大学轧制技术及连轧自动化国家重点实验室,辽宁沈阳110819 [2]广西先进铝加工创新中心有限责任公司,广西南宁530007 [3]广西南南铝加工有限公司,广西南宁530000

出　　处：《轻合金加工技术》2023年第1期23-29,58,共8页Light Alloy Fabrication Technology

基　　金：南宁市创新创业领军人才“邕江计划”资助项目(2019002);南宁市科技重大专项项目(20201041)。

摘　　要：在高端铝合金板带材生产过程中,准确、快速地识别和分析铝带材表面缺陷是控制带材表面质量的基础,也是现代智能生产过程智能感知的重要基础。针对气垫式连续热处理及表面处理生产线的高端铝合金汽车板、航空板的表面质量缺陷问题,采用融入协调注意力的MobileNetV2模型进行缺陷图像识别,提高模型对铝带材缺陷图像特征提取能力;融入协调注意力的MobileNetV2模型的准确率高达94.97%,准确率比原始MobileNetV2模型高1.34%,效果比肩ResNet50、VGG16模型;融入协调注意力的MobileNetV2模型的参数量仅1.52 MB,远远低于ResNet50、VGG16模型的参数量;该模型对缺陷识别精度高、识别速度快,具有很好的应用价值。In the production process of high-end aluminum alloy strip,accurate and rapid identification and analysis of aluminum strip surface defects is the basis of strip surface quality control,and also an important basis of intelligent perception in modern intelligent production process.Aiming at the surface quality defects of high-end aluminum alloy automobile sheet and aviation plate in the production line of air cushion continuous heat treatment and surface treatment,the MobileNetV2 model integrated with coordinated attention was used for defect image recognition,and the feature extraction ability of the model for aluminum strip defect image was improved.The accuracy of MobileNetV2 model integrated with coordinated attention is as high as 94.97%,1.34%higher than the original MobileNetV2 model.The effect is comparable to ResNet50 and VGG16 models.The number of parameters of Mobile-NetV2 model integrated with coordinated attention is only 1.52 MB,which is much lower than that of ResNet50 and VGG16 models.This model has good application value for high precision and fast identification speed of defects.

关 键 词：铝带材 表面缺陷 融入协调注意力 MobileNetV2模型 图像识别 深度学习 

分 类 号：TG146.21[一般工业技术—材料科学与工程] TP391.41[金属学及工艺—金属材料]