软件导刊2025，Vol.24Issue(10)：34-39,6.DOI:10.11907/rjdk.251192
基于特征改进的自监督工业图像异常检测
Self-Supervised Anomaly Detection in Industrial Images Based on Feature Improvement
摘要
Abstract
The anomaly detection of industrial images requires the recognition and localization of image defects.Some anomaly detection algo-rithms use pre trained models obtained from the ImageNet dataset for defect recognition,without paying attention to the impact of domain bias between ImageNet and industrial images on anomaly detection tasks.Therefore,a method is proposed to optimize the features of the memory bank to improve anomaly detection performance.Firstly,alleviate domain bias issues through coarse-grained segmentation training;Second-ly,guide the feature learning to learn the key information of the corresponding region block;Finally,improving the loss function enhances the clustering of similar sample features in the feature space,widens the class spacing between different sample features,and obtains better memo-ry features.On the MVTec AD dataset,the proposed method achieved 99.2%and 97.6%in AUROC metrics at the image and pixel levels,re-spectively.Compared to the benchmark model,it improved by 0.7%and 1.4%,indicating better detection performance.关键词
迁移学习/自监督学习/记忆库特征/粗粒度分割训练/工业图像异常检测Key words
transfer learning/self-supervised learning/memory bank features/coarse-grained segmentation training/industrial image anomaly detection分类
计算机与自动化引用本文复制引用
许昌源,贾可,金治成,李涵鑫,王文润,周记..基于特征改进的自监督工业图像异常检测[J].软件导刊,2025,24(10):34-39,6.基金项目
四川省科技计划项目(2023YFG0305) （2023YFG0305）
成都信息工程大学科研基金项目(KYTZ202156) （KYTZ202156）
