基于双通道卷积神经网络的缺陷检测  

Defect detection based on double channel convolutional neural network method

作　　者：张翠[1] 杨志清[1] 周茂杰[2] ZHANG Cui;YANG Zhi-qing;ZHOU Mao-jie(Bowen College of Management,Guilin University of Technology,Guilin,Guangxi 541006,China;Guilin University of Technology,Guilin,Guangxi 541004,China)

机构地区：[1]桂林理工大学博文管理学院,广西桂林541006 [2]桂林理工大学,广西桂林541004

出　　处：《模具工业》2020年第11期13-17,共5页Die & Mould Industry

基　　金：广西中青年教师基础能力提升项目(2018KY0250);广西科技计划项目(桂科AB17195028);桂林市科学技术研究开发项目(20170220)。

摘　　要：利用深度学习方法进行塑料产品缺陷检测时,存在训练样本不足,网络层次增加产生的梯度消失和梯度爆炸等问题,采用残差网络ResNet和压缩与激励网络SENet相结合,构建深度学习模型,解决梯度消失、梯度爆炸和注意力分布的问题,利用工业生产中的产品图像进行缺陷检测试验,经过2种试验结果分析,该算法有效提高了产品缺陷检测的准确率和召回率。In the process of plastic product defect detection by using deep learning method,there were some problems,such as insufficient training samples,gradient disappearance and explosion caused by increasing network level,etc..Combined ResNet with SENet to build a deep learning model,it solved the problems of gradient disappearance and explosion,and attention distribution.Images in industrial production were used to conduct defect detection experiments.Through the analysis of two experimental results,the algorithm could effectively improve the accuracy rate and recall rate of product defect detection.

关 键 词：ResNet SENet 缺陷检测 塑料产品 深度学习 

分 类 号：TG76[金属学及工艺—刀具与模具]