基于ResNet和DenseNet模型的铝罐表面涂层缺陷检测方法  

A Surface Coating Defect Detection Method for Aluminum Cans Based on ResNet and DenseNet Models

作　　者：谢凌望 陈满 彭宇瑞 杨海林 张兴伟[2] XIE Lingwang;CHEN Man;PENG Yurui;YANG Hailin;ZHANG Xingwei(Shantou Polytechnic,Shantou 515078,Guangdong,China;Shantou University,Shantou 515063,Guangdong,China;Shantou Oriental Technology Co.Ltd,Shantou 515144,Guangdong,China)

机构地区：[1]汕头职业技术学院,广东汕头515078 [2]汕头大学工学院,广东汕头515063 [3]汕头市东方科技有限公司,广东汕头515144

出　　处：《汕头大学学报（自然科学版）》2024年第4期40-48,共9页Journal of Shantou University：Natural Science Edition

基　　金：广东省科技计划项目(2020ST001);广东省博士工作站建设项目。

摘　　要：为了解决现有铝罐表面涂层缺陷检测难、检测精度低、漏检错检率高等问题,本文提出了一种基于卷积神经网络的铝罐表面涂层缺陷检测方法,通过对AlexNet、GoogLeNet、ResNet和优化后的DenseNet四种卷积神经网络模型进行训练,以选择最适合的检测模型实现对铝罐内、外表面涂层的缺陷检测,从而设计开发了用于铝罐自动化产线上的检测软件.结果表明,使用优化后的DenseNet18模型检测铝罐内表面涂层缺陷的准确率最高,为96.67%,使用ResNet18模型检测铝罐外表面涂层缺陷的准确率最高,为93.33%,两种模型的检测用时约为250 ms,其检测准确率和速度基本满足工业需求,在工业生产中达到了较好的实际应用效果.In order to solve the problems of difficult detection,low detection accuracy,and high leakage rate in the surface coating defect detection of aluminum cans,the convolutional neural networks is investigated to detect the surface coating defect of aluminum cans.By training four convolutional neural network models,AlexNet,GoogLeNet,ReseNet and optimized DenseNet,the most suitable model was selected to detect the inner and outer surface coating defects of aluminum cans.The detection software was designed and developed for automated aluminum can production lines.The result shows that the optimized DenseNet18 model has the highest accuracy in detecting the inner surface coating defects,at 96.67%.The ResNet18 model has the highest accuracy in detecting the outer surface coating defects,at 93.33%.The detection time of the two models is about 250 ms.The detection accuracy and speed of this method canmeet industrial requirementsbasically.This defect detection method has achieved good practical application results in industrial production.

关 键 词：铝罐 卷积神经网络 缺陷检测 

分 类 号：TU43[建筑科学—岩土工程] O344[建筑科学—土工工程]