基于异常生成的图片异常检测轻量化研究

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2026）03-0112-05

Research on Lightweight Image Anomaly Detection Based on Anomaly Generation

WANG Jie，YANHua （College ofElectronicsand Information Engineering，Sichuan University， Chengdu 61o065，China）

Abstract： Industrial image anomalydetection aims to identifyand locate defects onthesurfaces of industrial products. Existinganomalydetectionmethodsusuallhavelong inferencetimesandlarge modelparameters，making itdificult toachieve fastdetectiononresource-constrainededgedevices.Tosolvethisproblem，thispaperproposesalightweightanomalydetection model basedonanomaly generation.Bygenerating pixel-level defectpseudo-labels onimages and featuremaps，thetraditional unsupervisedanomalydetectiontaskis transformed intoasupervised pixelsegmentationtask，whichsignificantlyimproves detectionperformanceanddrasticallyreducescomputationaloverheadThe modelsructure issimple.Itfirstusesapre-trained residual network toextract features，thenperforms feature adaptationand featuredetail enhancement，and finalloutputsan anomalydistribution map through a segmentation head.Experimentsonthe MVTec AD dataset show that the model achieves high detection accuracy with fewer parameters.

Keywords： anomaly detection; anomaly generation; pixel segmentation; lightweight

0 引言

工业图片异常检测任务是识别出异常图片并定位异常区域。（剩余8183字）