基于改进领域对抗网络的瓷砖表面缺陷检测  

Ceramic tile surface defect detection based on improved domain-adversarial neural network

作　　者：林行 陈新度[1] 吴磊[1] 练洋奇 Lin Xing;Chen Xindu;Wu Lei;Lian Yangqi(College of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广州510006

出　　处：《电子测量技术》2022年第24期105-110,共6页Electronic Measurement Technology

基　　金：佛山市产业领域科技攻关项目(2020001006297);佛山市顺德区核心技术攻关项目(2030218000174)资助

摘　　要：深度神经网络作为主流的表面缺陷检测方法之一,需要大量样本进行模型训练,而随着瓷砖产品多样化,同类型瓷砖缺陷样本有限。本文提出一种基于改进域对抗神经网络(MDANN)的瓷砖表面缺陷检测方法,参考传统的DANN结构,首先,在ImageNet公共数据集上预训练保存网络参数,提高训练速度;然后,在原网络中加入瓶颈层,并利用最大均值差异指标优化领域分布差异,改善了原DANN网络筛选源域的能力,实现小样本瓷砖的缺陷检测。实验结果表明,MDANN对瓷砖表面缺陷的有效检出率达98.77%,相比于原DANN网络提高了3.53%,可快速适用于不同类型的瓷砖检测,泛化性好。Deep neural network is one of the mainstream surface defect detection methods,a large number of samples are needed for model training,but the defect samples of the same type of ceramic tile are limited with the diversification of ceramic tile products.In this paper,a ceramic tile surface defect detection method based on improved domain countermeasure neural network(MDANN)is proposed.Referring to the traditional DANN structure,the network parameters are pretrained on the ImageNet to improve the training speed.Then,the bottleneck layer is added to the original network,and the maximum mean difference index is used to optimize the field distribution difference,which improves the ability of the original DANN network to screen the source domain and realizes the defect detection of small sample tiles.The experimental results show that the effective detection rate of MDANN for ceramic tile surface defects achieves 98.77%,which is 3.53%higher than the original DANN network.It can be quickly applied to the detection of different types of ceramic tiles with good generalization.

关 键 词：瓷砖缺陷检测 深度学习 迁移学习 领域自适应神经网络(DANN) 

分 类 号：TQ174.765[化学工程—陶瓷工业] TP183[化学工程—硅酸盐工业] TP391.41[自动化与计算机技术—控制理论与控制工程]