组会ppt-使用并改进扩散模型用于工业检测[ECCV2024]GLAD Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection.pptx

下载文档

0
0
约3.65千字
约 18页
2024-11-27 发布于江苏
举报
版权申诉
保障服务

组会ppt-使用并改进扩散模型用于工业检测[ECCV2024]GLAD Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection.pptx

1、本文档共18页，可阅读全部内容。
2、有哪些信誉好的足球投注网站（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

ECCV2024

BackgroundAnomalydetection(AD)aimstodetectandlocateabnormalpatternsofobjects:itchallengingtocollectenoughabnormalsamplesforallanomalytypesinsituations.ever-changingproductdesignandproductionprocesses,itisimpossibletocollectallanomaliesinadvance.unsupervisedanomalydetection(UAD)hasdrawnmuchattentionwithonlynormalsamplesrequired.

RelatedworksEmbedding-basedmethods:extractfeatureofimagestoevaluateabnormal.Knowledgedistillation-basedmethods:trainstudentnetworkwithnormalsamples,featuresfromthepre-trainedteachernetworkarecomparedwithfeaturesfromthestudentnetworktodetectandlocateanomalies.PaDiMbuildsmultivariateGaussiandistributionsforpatchfeaturesofnormalsamplesandusesMahalanobisdistanceastheanomalyscore.PatchCoreproposesamemorybanktosavefeaturesofnormalimages,whicharecomparedwithfeaturemapsoftestimagestodistinguishthedifferencebetweennormalandabnormalfeatures.Reconstruction-basedmethods:detectedandlocatedviathecomparisonbetweenthegivensampleanditsnormalcounterpart.Basedonthehypothesisthatmodelstrainedonnormalsamplesonlycanreconstructnormalimageswell.Anomaliescanbedetectedbycomparingthesamplesbeforeandafterreconstruction.AE(early),GAN,transformer,UNetarchitecture

MotivationDiffusionmodelshaveprominentmodelingability.Duringthetrainingprocess,thediffusionmodelcapturesthedistributionofnormalsamplesonly.Thesamedenoisingstep:Differentanomaliesisuneven.LesspreserveddetailsoftheoriginalTheanomalynoiseinevitablydeviatesfromthestandardGaussiandistribution.?

MethodInference:AdaptiveDenoisingStep(ADS):achievesabettertrade-offbetweenreconstructionqualityanddetailpreservationability.Spatial-AdaptiveFeatureFusion(SAFF):avoidreconstructionofnormalregions.Training:Anomaly-orientedTrainingParadigm(ATP):allowdiffusionmodeltopredictno