基于融合特征分布学习与图像重建的异常检测  

Abnormity detection based on fusion feature distribution learning and image reconstruction

作　　者：朱思宇 朱磊[1] 王文武[1] 乐华钢 ZHU Siyu;ZHU Lei;WANG Wenwu;YUE Huagang(School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430070,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430070

出　　处：《液晶与显示》2024年第8期1116-1127,共12页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.62173262)。

摘　　要：无监督学习是当前工业产品缺陷检测领域的主流研究方向,目前主要分为基于重建和基于特征的两类方法。前者构建基于内容感知的映射以将异常区域映射为正常区域并通过残差图像来检测缺陷,注重于图像整体的表现。后者利用高层语义特征以实现定位异常,更加注重图像细节呈现。根据两种方法的优缺点,本文提出一种基于特征与重建融合的方法,有效结合两者优点互补其不足并实现统一的端到端的学习与推理。首先训练一个重建模型,然后采用归一化流模型以充分学习输入正常样本的高可能性数据概率分布,使其与重建模型相融合,有效地提高重建模型缺陷检测以及缺陷定位的准确率。在广泛使用的MVTec AD数据集上,提出的融合模型的平均图像级AUROC达到了98.7%,平均像素级AUROC达到了94.2%,特别地,相比单一重建模型提升了3.3%。提出的特征与重建网络融合模型显著提高了重建网络部分对于缺陷定位的不足,使结果更为精确。Unsupervised learning is the main research direction in the field of industrial product defect detection at present,and it is mainly divided into two types of methods:reconstruction based and feature based methods.The former constructs content-aware mappings to map abnormal regions into normal regions and detect defects through residual images,focusing on overall performance of the images.The latter uses high-level semantic features to achieve positioning exceptions and pay more attention to image detail presentation.According to the advantages and disadvantages of the two methods,a method is proposed based on the fusion of characteristics and reconstruction,which effectively combines advantages of the two methods to complement their shortcomings and realize unified end-to-end learning and reasoning.A reconstructed model is trained firstly,then a normalized flow model is adopted to fully learn the probability distribution of high probability data of input normal samples,and it is integrated with the reconstructed model to effectively improve the accuracy of defect detection and defect positioning of the reconstructed model.On the widely used MVTec AD data set,the average image level AUROC of the proposed fusion model reaches 98.7%,the average pixel-level AUROC reaches 94.2%,in particular,an increase of 3.3%compared to a single reconstruction model.The convergence model of characteristics and proposed reconstruction network has significantly improved the shortcomings of defect positioning in the reconstruction network part,which makes experimental results more accurate.

关 键 词：异常检测 重建 内存模块 流模型 融合算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]