集成重构与分割网络的工业异常无监督检测算法
打开文本图片集
关键词：异常检测；无监督；多尺度特征；动态界限界定信息；耦合注意力中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2025）12-035-3799-08doi：10.19734/j.issn.1001-3695.2025.03.0146
Unsupervised industrial anomaly detection via integrated reconstruction and segmentation networks
Wang Ruochen，Zhou Yi† （SchoolofArtificialIntellgenceandAutomation，WuhanUniversityofScienceandTechnology，Wuhan43O81China）
Abstract：Unsupervisedanomalydetectionin industrialproduct inspectionpresentsasignificantchallenge forcomputervision systems，with syntheticsample-based approaches emerging asanactiveresearcharea.To address thechaenges including the lackofcross-hierarchicalresponseoptimization，theinabilityofstaticconstraintmechanisms toadapt toojectboundaryvariations，andthecouplingoffeaturesinthelatent spaceleading todificulties inestablishingcompactdistribution boundaries， this studyproposedaboundary-awareguidanceintegrationnetwork（BGIN）framework integratingreconstructionandsegmentationnetworks.Themethodologyestablishedthree technicalcontributions：firstly，theframeworkemployedahierarchicalfeaturepyramidarchitecturewithintheintegratednetwork toachieveulti-salefeatureco-optimization，nhancingcro-solution semanticconsistency.Secondly，thestudydesignedadynamicboundary-awareguidance mechanism tofacilitatecrosmodal interactionbetweenthereconstructionandsegmentationbranches.Thisenabledcolaborativeoptimizationbetweenthereconstructionandsegmentationnetworks.Thirdly，thenetwork’sdiscrepancy-sensitivefeaturecouplingarchitectureintegratedparalelconvolutionallayers withadual-pathcoupledatentionmodule.Thiscomponentamplifiedfeatureresponses inanomalous regionsthrough inter-channel dependency modeling.Experimentalvalidationonthe MVTecADbenchmark demonstrates the framework’s superior performance，achieving state-of-the-art image-level and pixel-level AUC scoresof 99.2% and 97.8% ， respectively. Comparative analysis reveals 55.5% false detection reduction against the leading synthetic method，JRCC-Net. Thesolutionefectivelyhandlescomplextexturesandmicro-defectscenariosthroughitshierarchicalfeature integrationmechanism.
ey Words：anomalydetection;unsupervised；multi-scale features;dynamic object boundary constraints；coupledattel
0 引言
工业产品质量检测是智能制造领域的关键环节，直接关乎生产安全与经济效益[1]。（剩余21352字）
