电子学报2025，Vol.53Issue(3)：895-909,15.DOI:10.12263/DZXB.20240733
基于双向约束蒸馏的无监督图像异常检测
Unsupervised Image Anomaly Detection Based on Constrained BidiRectional Distillation
摘要
Abstract
Anomaly detection has been widely studied and applied to various visual scenes.Recently,the mainstream unsupervised anomaly detection schemes are usually based on distillation methods and reconstruction methods.However,they still have some limitations.In distillation model,the student network can usually learn the strong representation ability of the teacher network,thus can not represent differently for the abnormal regions.In reconstruction model,the encoder-de-coder model can easily learn a restoration shortcut and recover features indiscriminately.To address the above challenges,we propose N-Net,which integrates the advantages of above two methods and alleviates limitations through the bidirection-al distillation module and the multistage filtration mechanism.Specifically,in the teacher-student network,this paper first proposes distilling adaptive domain features instead of original domain features,which ensures efficient alignment of nor-mal adaptive domain features through bidirectional distillation branches.Then,we propose a multilevel filtering module to filter abnormal features through query and compression to further enhance the ability to learn normal semantic feature distri-bution and improve the anomaly detection performance.Finally,a large number of experiments are carried out on two benchmark anomaly detection datasets,MVTec and VisA.The results show that the proposed method achieves advanced performance in anomaly detection and location tasks.关键词
异常检测/双向蒸馏/特征映射/多级过滤/特征压缩Key words
anomaly detection/bidirectional distillation/feature projection/multistage filtration/feature compact分类
信息技术与安全科学引用本文复制引用
李波,李泽超,邢鹏,唐金辉..基于双向约束蒸馏的无监督图像异常检测[J].电子学报,2025,53(3):895-909,15.基金项目
国家自然科学基金(No.U20B2064,No.U21B2043) National Natural Science Foundation of China(No.U20B2064,No.U21B2043) （No.U20B2064,No.U21B2043）
