Automatic Localization of Casting Defects withConvolutional Neural Networks 铸件缺陷的自动定位 卷积神经网络

《Automatic Localization of Casting Defects withConvolutional Neural Networks 铸件缺陷的自动定位 卷积神经网络》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、AutomaticLocalizationofCastingDefectswithConvolutionalNeuralNetworksMaxFergusonRonayAkYung-TsunTinaLeeKinchoH.LawEngineeringInformaticsGroupSystemsIntegrationDivisionSystemsIntegrationDivisionEngineeringInformaticsGroupCivilandEnvironmentalNationalInsti

2、tuteofStandardsNationalInstituteofStandardsCivilandEnvironmentalEngineeringandTechnology(NIST)andTechnology(NIST)EngineeringStanfordUniversityGaithersburg,UnitedStatesGaithersburg,UnitedStatesStanfordUniversityStanford,UnitedStatesronay.ak@nist.govyung-

3、tsun.lee@nist.govStanford,UnitedStatesmaxferg@stanford.edulaw@stanford.eduAbstract—AutomaticlocalizationofdefectsinmetalcastingsisThereareanumberofnondestructiveexamination(NDE)achallengingtask,owingtotherareoccurrenceandvariationintechniquesavailablefo

4、rproducingtwo-dimensionalandthree-appearanceofdefects.Convolutionalneuralnetworks(CNN)havedimensionalimagesofanobject.Real-timeX-rayimagingrecentlyshownoutstandingperformanceinbothimagetechnologyiswidelyusedindefectdetectionsystemsinclassificationandloc

5、alizationtasks.Weexaminehowseveralindustry,suchason-linewelddefectinspection[3].UltrasonicdifferentCNNarchitecturescanbeusedtolocalizecastingdefectsinspectionandmagneticparticleinspectioncanalsobeusedtoinX-rayimages.Wetakeadvantageoftransferlearningtoal

6、lowmeasurethesizeandpositionofcastingdefectsincaststate-of-the-artCNNlocalizationmodelstobetrainedonacomponents[4,5].Analternativemethodisthree-dimensionalrelativelysmalldataset.Inanalternativeapproach,wetrainaX-raycomputedtomography,thatcanbeusedtovisu

7、alizethedefectclassificationmodelonaseriesofdefectimagesandtheninternalstructureofmaterials.RecentdevelopmentsinhighuseaslidingclassifiermethodtodevelopasimplelocalizationresolutionX-raycomputedtomographyhavemadeitpossibletomodel.Wecomparethelocalizatio

8、naccuracyandcomputationalgainathree-dimensionalcharacterizationofporosity[6,7].performanceofeachtechnique.WeshowpromisingresultsfordefectlocalizationontheGRIMAdatabaseofX-rayimagesThedefectdetectionprocesscanbeframedaseitheran(GD