使用 FOMO-AD 进行视觉异常检测.pdf

《使用 FOMO-AD 进行视觉异常检测.pdf》由会员分享，可在线阅读，更多相关《使用 FOMO-AD 进行视觉异常检测.pdf（25页珍藏版）》请在三个皮匠报告上搜索。

1、Visual Anomaly Detection with FOMO-ADJan JongboomCo-founder&CTOEdge ImpulseLeading development platform for machine learning on edge devices103,933 new projects(!)created since last Embedded Vision Summit40%of these are vision projectsEdge Impulse2Edge Impulse project countCan you trust ML models?3h

2、ttps:/ with an unknown state4GiraffeZebraOtherDataset asymmetry5Normal operationFaulty operationDataset asymmetry6Fault state 1Fault state 2Fault state 3Fault state 4Normal operationAnomaly detection7Training dataAll potentialinputsxUnseen input(no anomaly)xUnseen input(anomaly)Auto-encoders?8Input

3、imageReconstructed imagediff w/input image:similar?no anomaly.Computationally expensive,need both encoder/decoder.Working in pixel space is not great:poor evaluation metric,blurry images.Visual anomaly detection requires very high resolution images.Same accuracy:106parameters(auto-encoder)vs 103para

4、meters(our new approach).Why auto-encoders dont work9Anomaly detection on sensor data10DSP FeaturesInputClustering algorithmGreat for basic sensor datafor which you can reason about featuresClustering with Gaussian Mixture Models11Applying this to visual AD12InputNeural network embeddingsGaussian Mi

5、xture ModelsOperates on feature space,not pixel spaceTesting out this premise:MobileNet13Testing out this premise:MobileNet14In lots of visual inspection cases only part of the image is anomalous.Input image might be very large,if only 0.5%of your image covers an anomaly=hard to get your loss functi

6、on right.Knowing where a fault is,is super useful for humans!Where is the anomaly?15Edge Impulse FOMO16Each cell is aclassifierEdge Impulse FOMO17Each cell is aclassifierReplace classification with a GMM per cell.Each cell now has an anomaly score.Fully convol