quotation:[Copy]
[Copy]
【Print page】 【Download 【PDF Full text】 【View/Add Comment】【Download reader】【 Close 】

←Previous page|Page Next →

Back Issue    Advanced search

This Paper:Browse 232   Download 214 本文二维码信息
码上扫一扫!
基于门控注意力的知识蒸馏异常检测方法
江结林,刘丁豪,刘西应,崔燕
0
(1.南京信息工程大学 软件学院,南京 210044;2.南京特殊教育师范学院 数学与信息科学学院,南京 210038)
摘要:
针对知识蒸馏在无监督异常检测任务中逻辑异常检测能力不足的问题,提出了一种基于门控注意力的知识蒸馏网络。该网络的思想是通过引入异常特征限制模块对学生网络的异常特征进行多层次限制,使其专注于正常特征的学习,并利用特征压缩模块动态调整全局特征权重,以及通过多尺度特征聚合模块联合浅层与深层特征,同时动态调整局部特征权重,从而在保证结构异常检测的基础上提升逻辑异常的检测性能。实验结果表明,该方法在基准数据集MVTec AD的像素级平均准确率达到98.52%。
关键词:  异常检测  知识蒸馏  门控注意力  逻辑异常
DOI:10.20079/j.issn.1001-893x.250214001
基金项目:国家自然科学基金资助项目(62001236)
A Gated Attention-based Method for Knowledge Distillation Anomaly Detection
JIANG Jielin,LIU Dinghao,LIU Xiying,CUI Yan
(1.School of Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;2.College of Mathematics and Information Science,Nanjing Normal University of Special Education,Nanjing 210038,China)
Abstract:
In order to solve the problem of insufficient logical anomaly detection ability in the unsupervised anomaly detection task of knowledge distillation,a gated attention-based knowledge distillation network is proposed.The idea of the network is to introduce the anomalous feature restriction module to restrict the abnormal features of the student network at multiple levels,so that it can focus on the learning of normal features,and use the feature compression module to dynamically adjust the global feature weight,and through the multi-scale feature aggregation module to combine the shallow and deep features,and dynamically adjust the local feature weight,so as to improve the detection performance of logical anomalies on the basis of ensuring structural anomaly detection.Experimental results show that the average pixel-level accuracy of the proposed method in the benchmark dataset MVTec AD reaches 98.52%.
Key words:  anomaly detection  knowledge distillation  gated attention  logical anomalies