改进GANomaly模型的印刷辊筒表面缺陷检测方法  

Detection Method of Printing Roller Surface Defects Based on Improved Ganomaly Model

作　　者：王振强 杜昊晨 彭泽 WANG Zhenqiang;DU Haochen;PENG Ze(Qingdao Broadcasting Comprehensive Information Center,Qingdao Shandong 266071,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China;Taicang Institute of Information Technology,Taicang Jiangsu 215400,China)

机构地区：[1]青岛广播电视综合信息中心有限公司,山东青岛266071 [2]华北电力大学控制与计算机工程学院,北京102206 [3]太仓中科信息技术研究院,江苏太仓215400

出　　处：《信息与电脑》2022年第23期190-193,256,共5页Information & Computer

摘　　要：工业产品中的缺陷样本获取困难,而且缺陷的表现形式多种多样,因此常使用无监督异常检测方法检测工业产品表面缺陷。无监督异常检测通过计算异常分数来判断图像中是否包含缺陷,但是如何准确定位缺陷是亟待解决的问题。为了解决该问题,提出了一种基于改进的GANomaly模型异常检测方法。该方法首先在原来模型的基础上加入了异常检测模块获取异常分数,其次对利用梯度和最大信息熵的图片,使用分水岭分割算法和特征对齐对缺陷进行定位,最后使用E-measure评估分割结果。实验结果表明,设计方法在印刷辊筒数据集上的检测与分割效果均优于其他无监督异常检测方法。It is difficult to obtain the sample of defects in industrial products,and the manifestations of defects are diverse.unsupervised anomaly detection methods are widely used in the field of industrial product surface defect detection.Unsupervised anomaly detection is to determine whether the image contains defects by calculating the anomaly score,but how to accurately locate defects is an urgent problem to be solved.In order to solve this problem,this paper proposes an anomaly detection method based on the improved GANomaly model.Based on the original model,the method adds an anomaly detection module to obtain the anomaly score,then uses gradient and maximum information entropy,watershed segmentation algorithm and feature alignment to locate defects,and finally uses E-Measure to evaluate the segmentation results.The experimental results show that the detection and segmentation effects of this method on the printing roller dataset are better than other unsupervised anomaly detection methods.

关 键 词：异常检测 特征对齐 分水岭算法 缺陷检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]