面向工件表面缺陷的无监督域适应方法  

Unsupervised Domain Adaptation for Surface Defects of Industrial Component

作　　者：姚明海[1] 沈颖琦 柳嘉敏 YAO Ming-hai;SHEN Ying-qi;LIU Jia-min(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310000,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310000

出　　处：《小型微型计算机系统》2021年第6期1256-1261,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61871350)资助.

摘　　要：工件表面缺陷检测是保障工件质量的重要环节,在有足够标签的情况下,有监督学习可以很好地对其缺陷进行分类,但当有新的工件需要检测时,又要标注新的数据集.为了解决该问题,本文利用不同工件之间缺陷种类的相似性,并基于迁移学习的思想,提出了一种全局特征和局部特征共同适应的无监督域适应方法(GLDA).该方法首先利用生成对抗网络对齐全局特征,然后最小化生成域数据与目标域数据的协方差矩阵的差从而对齐局部特征,并将语义损失加入总体优化目标函数中作为正则项,使得对齐全局特征后的数据经过分类网络后不丢失局部的特征信息.所提出的方法不再局限于同类别不同分布的迁移,并在工件表面数据集和公开数据集上能取得较好的分类效果.Defect detection on the surface of industrial component is an important part of quality assurance.Supervised learning can classify its defects well with labeled data,but when there need a new production to detect,this product must be labeled.In order to solve this problem,basing on the similarity of defect types between different industrial component and transfer learning method,we proposed an unsupervised method(GLDA)that adapts global feature and local feature.The method firstly uses the generative adversarial networks to align global features,then minimizes the difference between the covariance matrix of generated domain data and target domain data to align local features,and it also adds a semantic loss to the overall optimization goal as a regular term,which can ensure local features retained when align global features.The proposed method is not limited to the different distributions in same category,and it achieves a good classification result on the industrial component surface dataset and the public dataset.

关 键 词：缺陷检测 无监督学习 全局特征 局部特征 域适应 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]