压敏电阻表面缺陷深度无监督检测方法  

Deep unsupervised multi-color and multi-specification varistor surface defect detection

作　　者：唐善成[1] 张莹 陈明 张雪 TANG Shancheng;ZHANG Ying;CHEN Ming;ZHANG Xue(School of Communication and Information Engineering,Xi'an University of Science and Technology,Xi'an 710600,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710600

出　　处：《华中科技大学学报（自然科学版）》2023年第9期152-159,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

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(2019ZDLSF07-06);西安市科技计划科技创新引导项目(201805036YD14CG20(4));陕西省科技计划重点产业创新链(群)-工业领域项目(2020ZDLGY15-07).

摘　　要：针对表面缺陷检测方法泛化能力不足问题,提出一种深度迁移卷积变分自编码网络进行压敏电阻表面缺陷检测.检测方法包含多颜色多规格压敏电阻图像自适应预处理、模型建立及缺陷检测3个阶段.第一阶段完成自适应规范化、自适应通道选取、自适应阈值处理;第二阶段完成参数迁移,自动提取良品差分图像特征得到模型;第三阶段对差分图像与重构图像进行差分处理后得到的残差图像,通过改进的“3σ准则”进行缺陷检测.结果表明:方法平均检测准确率达98.83%,相比其他流行模型至少提高了9.53%,检测范围扩展至2种颜色3种规格的压敏电阻.Aiming at the problem of insufficient generalization ability of surface defect detection methods,a deep transfer convolutional variational autoEncoder(DTCVAE)was proposed for varistor surface defect detection.The method includes multicolor and multi-specification varistor image(MMVI)adaptive preprocessing,DTCAVE model establishment and defect detection 3 stages.The first stage completes adaptive normalization,adaptive channel selection,and adaptive threshold processing;the second stage completes parameter migration,and automatically extracts the characteristics of the difference image of the good product to obtain the model;the third stage detects the residual image through the improved"3σcriteria".The results show that the average detection accuracy of the method is 98.83%,which is at least 9.53%higher than other popular models.The detection range is extended to 2 colors and 3 specifications of varistors.

关 键 词：缺陷检测 压敏电阻 变分自编码器 参数迁移 无监督学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]