[IJCNN 2012] Steel Defect Classification with Max-Pooling Convolutional Neural Networks

《[IJCNN 2012] Steel Defect Classification with Max-Pooling Convolutional Neural Networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、SteelDefectClassificationwithMax-PoolingConvolutionalNeuralNetworksJonathanMasci,UeliMeier,DanCiresan,GabrielFricoutJurgenSchmidhuber¨ArcelorMittalIDSIA,USIandSUPSIMaizieresResearchSA,`Galleria2,6928Manno-Lugano,FranceSwitzerlandfgabriel.fricout@arcelormittal.comgfjona

2、than,ueli,dan,juergeng@idsia.chAbstract—WepresentaMax-PoolingConvolutionalNeuralsification.Thesystemisusuallybasedonasetofhand-Networkapproachforsupervisedsteeldefectclassification.Onawiredpipelineswithpartialornoself-adjustableparametersclassificationtaskwith7defects,co

3、llectedfromarealproductionwhichmakesthefine-tuningprocessofthisindustrialsystemsline,anerrorrateof7%isobtained.ComparedtoSVMcumbersome,requiringmuchmorehumaninterventionthanclassifierstrainedoncommonlyusedfeaturedescriptorsourbestnetperformsatleasttwotimesbetter.Notonly

4、wedoobtaindesired.Inthisworkwefocusonthetwolastpipelinestagesmuchbetterresults,buttheproposedmethodalsoworksdirectlyandproposeanapproachbasedonMax-PoolingConvolutionalonrawpixelintensitiesofdetectedandsegmentedsteeldefects,NeuralNetworks(MPCNN)[1],[2],[3],[4],[5],that

5、learnavoidingfurthertimeconsumingandhardtooptimizead-hocthefeaturesdirectlyfromlabeledimagesusingsupervisedpreprocessing.learning.Weshowthattheproposedmethodachievesstate-of-the-artresultsonrealworlddataandcompareourapproachI.INTRODUCTIONtoclassifierstrainedonclassicfe

6、aturedescriptors.MachinevisionbasedsurfaceinspectiontechnologieshaveThereisnotmuchliteratureaboutsteeldefectdetectiongainedalotofinterestfromvariousindustriestoautomatein-[6].However,inabroadercontexttheproblemcanbespectionsystems,andtosignificantlyimproveoverallproduc

7、tviewedasdefectdetectionintexturedmaterialwhichhasquality.Atypicalindustryadoptingtheserefinedinspectionreceivedconsiderableattentionincomputervision[7],[8],toolsistherolledsteelstripmarket.Real-timevisualinspec-[9].Inclassicalapproaches,featureextractionisperformedtio

8、nofproductionlinesiscrucialtoprovideaproductwithusingthefilter-bankparadigm,resultinginanarchitectureeverfewersurfacedefects.