织物疵点检测的反向传播神经网络分类器优化设计的经验方法(IJISA-V6-N9-4)

I.J.IntelligentSystemsandApplications,2014,09,30-39PublishedOnlineAugust2014inMECS()DOI:10.5815/ijisa.2014.09.04Copyright©2014MECSI.J.IntelligentSystemsandApplications,2014,09,30-39AnEmpiricalMethodforOptimizationofCounterpropagationNeuralNetworkClassifierDesignforFabricDefectInspectionMd.TarekHabib,Mr.,Asst.Prof.DepartmentofComputerScienceandEngineering,GreenUniversityofBangladesh,Dhaka,BangladeshE-mail:[e-mail],Dr.,Prof.DepartmentofElectricalEngineeringandComputerScience,NorthSouthUniversity,Dhaka,BangladeshE-mail:[e-mail]—Automated,i.e.machinevisionbasedfabricdefectinspectionsystemshavebeendrawingplentyofattentionoftheresearchersinordertoreplacemanualinspection.Twodifficultproblemsaremainlyposedbyautomatedfabricdefectinspectionsystems.Theyaredefectdetectionanddefectclassification.Counterpropagationneuralnetwork(CPN)isarobustclassifierandverypromisingfordefectclassification.Ingeneral,worksreportedtodatehaveclaimedvaryinglevelofsuccessesindetectionandclassificationofdifferenttypesofdefectsthroughCPN;butinparticular,noclaimedhasbeenmadeforsuccessfulapplicationofCPNforfabricdefectsdetectionandclassification.Inthosepublishedworks,noinvestigationhasbeenreportedregardingtothevariationofmajorperformanceparametersofNNbasedclassifierssuchaslearningtimeandclassificationaccuracybasedonnetworktopologyandtrainingparameters.Asaresult,applicationengineerhaslittleornoguidancetotakedesigndecisionsforreachingtooptimumstructureofNNbaseddefectclassifiersingeneralandCPNbasedinparticular.OurworkfocusesonempiricalinvestigationofinterrelationshipbetweendesignparametersandperformanceofCPNbasedclassifierforfabricdefectclassification.ItisbelievedthatsuchworkwillbelayingthegroundtoempowerapplicationengineerstodecideaboutoptimumvaluesofdesignparametersforrealizingmostappropriateCPNbasedclassifier.IndexTerms—FabricDefect,MachineVision,DefectClassification,NeuralNetwork(NN),CounterpropagationNeuralNetwork,OptimizationProblem,OptimumDesignParameter.I.INTRODUCTIONProductqualityassuranceistreatedasoneofthemostsignificantfocusesintheindustrialproduction.Productqualityisseverelylessenedbydefects.Failuretoearlydefectdetectionincurscostsintermsoftime,moneyandconsumersatisfaction.So,earlyandaccuratedefectdetectionisanimportantaspectofqualitycontrol.Manualinspectionistimeconsumingandtheaccuracylevelisnotgoodenoughtomeetthepresentdemandofthehighlycompetitivenationalandinternationalmarket.Thesolutiontotheproblemsposedbymanualinspectionidautomated,i.e.machinevisionbaseddefectinspectionsystem.Thisiswhy,machinevisionbaseddefectinspectionsystemisverychallengingtopicforresearchinvariousdomainsofindustrialproducts,e.g.integratedcircuits,printedcircuitboards,ballgridarrays[1],ceramictiles[2],sandpaper,castings,leather[3]andevencigarettespackagedinatincontainer[4].Likewisemachinevisionbasedfabricdefectinspectionsystemisagoodthrustfortheresearchersofmanycountries.Automatedfabricdefectinspectionsystemsmainlyinvolvetwochallengingproblems,namelydefectdetectionanddefectclassification.Automatedfabricdefectinspectionsystemsarereal-timeapplications.Sotheyrequirereal-timecomputation,whichexceedsthecapabilityoftraditionalcomputing.Neuralnetworks(NNs)aresuitableenoughforreal-timesystemsbecauseoftheirparallel-processingcapability.Moreover,NNshavestrongcapabilitytohandleclassificationproblemswithgoodclassificationaccuracy.Theyvaryinnetworkarchitectureaswellastrainingorlearningalgorithm.ThereisanumberofperformancemetricsofNNmodels.Classificationaccuracy,modelcomplexityandtrainingtimearethreeofthemostimportantperformancemetricsofNNmodels.Counterpropagationneuralnetworks(CPNs)canhavegoodperformanceasclassifiers.Theycanbeemployedinreal-timesystems.Theyarehybridnetwork,whicharecapableofhandlingcomplexclassificationproblemswithgoodclassificationaccuracy[5-7].Again,thenumberofcomputingunitsinaCPNmodelislow.Thismakesnetworktopologysimple,i.e.modelcomplexitybecomeslow.Moreover,differenttypesoflearningalgorithmsareemployedforeachlayerinaCPN,whichresultsinshorttrainingtimeofthenetwork[7,8].SoaCPNappearstobeaverygoodchoiceasaclassifierinordertoaddresstheproblemoffabricdefectclassification.Althoughtherehavebeensomereportsaboutthefeasibilityofneuralnetworkbasedclassifierdevelopmentforfabricdefectclassification,buttherehasbeennoreportedworkinvestigatinginterrelationshipbetweendesignparametersandperformanceofNNbasedclassifier.ConceptdemonstrationaloneisnotsufficienttoempoweranapplicationengineertodesignoptimumAnEmpiricalMethodforOptimizationof31CounterpropagationNeuralNetworkClassifierDesignforFabricDefectInspectionCopyright©2014MECSI.J.IntelligentSystemsandApplications,2014,09,30-39classifier.Therefore,thisworknotonlyfocusesonthestudyofthefeasibilityofCPNmodelinthecontextoffabricdefectclassification,butalsoreportsthefindingsofempiricalinvestigationabouttheimplicationsofCPNdesignparametersonthetrainingandclassificationperformance.Inparticular,weempiricallydiscovertheinterrelationshipbetweentheperformancemetr