采用变分自编码器的无监督压敏电阻表面缺陷检测  

Unsupervised varistor surface defect detection based on variational autoencoder

作　　者：唐善成[1] 陈明 王瀚博 张雪 张莹 TANG Shancheng;CHEN Ming;WANG Hanbo;ZHANG Xue;ZHANG Ying(School of Communication and Information Engineering, Xi'an University of Science and Technology, Xi'an 710054,China)

机构地区：[1]西安科技大学通信与信息工程学院,陕西西安710054

出　　处：《计算机集成制造系统》2022年第5期1337-1351,共15页Computer Integrated Manufacturing Systems

基　　金：国家重点研发计划资助项目(2018YFC0808300);陕西省重点研发计划资助项目(2018GY-151,2019ZDLSF07-06);西安市科技计划资助项目(201805036YD14CG20(4))。

摘　　要：为解决在压敏电阻表面缺陷检测中缺陷样本采集困难以及检测精度差的问题,提出一种采用深度卷积变分自编码器(DCVAE)的无监督表面缺陷检测方法,用于压敏电阻表面缺陷的检测。预处理阶段利用去背景、规范化以及图像差分技术对原始图像进行处理,精准提取出了差分图像,消除了无关数据特征对算法的影响。为了提升缺陷检测的精度,算法首先将空间注意力机制融入变分自编码器,有效地提取出了良品图像的特征;其次将重构前后图像相减得到残差图,突出缺陷区域,衰减非缺陷区域。通过与已有流行的两种无监督缺陷检测方法进行对比,实验结果表明,所提方法F1值至少提升了7.4%,准确率达98.58%。To solve the problems of defect samples collection and poor detection accuracy in the surface defect detection of varistor,an unsupervised surface defect detection method using Deep Convolutional Variational Autoencoder(DCVAE)was proposed.At the pre-processing stage,the original images were processed by background elimination,normalization and image difference technology,and then the difference images were extracted accurately to eliminate the influence of irrelevant data features.To improve the accuracy of defect detection,the proposed algorithm integrated spatial attention mechanism into the variational autoencoder to effectively extract the features of good products.The residual image obtained by subtracting the reconstructed image and the original one was used to highlight the defect area and attenuated the nondefect area.Compared with two popular unsupervised defect detection methods,the experimental results showed that the F1 value of the proposed method was improved by at least 7.4%,and the accuracy rate was 98.58%.

关 键 词：缺陷检测 压敏电阻 无监督学习 空间注意力 变分自编码器 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]