基于改进Faster-RCNN的偏光片表面缺陷检测  

Surface defect detection of polarizer based on improved Faster-RCNN

作　　者：夏禹 肖金球[1,2] 翁玉尚 XIA Yu;XIAO Jinqiu;WENG Yushang(School of Electronics and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China;Intelligent Measurement and Control Engineering Technology Research Center,Suzhou 215009,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215009 [2]苏州市智能测控工程技术研究中心,江苏苏州215009

出　　处：《光学技术》2021年第6期695-702,共8页Optical Technique

基　　金：江苏省产学研前瞻性联合项目基金(BY2011132);江苏省研究生创新与教改项目(09150001)。

摘　　要：针对人工和传统自动化检测偏光片表面缺陷的准确性和效率问题,解决传统机器视觉在人工设计特征和泛化能力差的问题,提出了一种基于改进Faster-RCNN的偏光片表面缺陷检测方法。通过对比四种特征提取网络最终选择ResNet-101并引入特征金字塔网络(FPN)来提高对小缺陷的检测能力;接着采用ROI Align取代原始的ROI Pooling以解决两次取整操作引起的像素误差;最后通过采集方案获取偏光片表面图像,建立三种缺陷类型的数据集,结合k-means++聚类算法来改进anchor生成方案。实验表明,改进后的网络在偏光片缺陷数据集的mAP达到93.5%,平均检测单张待测图像耗时0.142s,能够满足实际检测的需求。Aiming at the accuracy and efficiency of manual and traditional automatic detection of polarizer surface defects,and solving the problem of poor manual design features and generalization capabilities of traditional machine vision,a polarizer surface defect detection method based on improved Faster-RCNN is proposed.First,by comparing the four feature extraction networks,finally select ResNet-101 and introduce the Feature Pyramid Network(FPN) to improve the detection ability of small defects;then use ROI Align instead of the original ROI Pooling to solve the pixel error caused by two rounding operations;Finally,the polarizer surface image is acquired through the acquisition scheme,and three types of defect data sets are established,and the k-means++ clustering algorithm is combined to improve the anchor generation scheme.Experiments show that the improved network has a mAP of 93.5% in the polarizer defect dataset,and the average detection time for a single image to be tested is 0.142 s,which can meet the needs of actual detection.

关 键 词：表面缺陷检测 卷积神经网络 FPN Faster-RCNN k-means++ 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]