Emergent self-organizing feature map for recognizi

ORIGINALARTICLEEmergentself-organizingfeaturemapforrecognizingroadsignimagesYok-YenNguwi•Siu-YeungChoReceived:28January2009/Accepted:15October2009/Publishedonline:3November2009Springer-VerlagLondonLimited2009AbstractRoadsignrecognitionsystemremainsachal-lengingpartofdesigninganIntelligentDrivingSupportSystem.Whilethereexistmanyapproachestoclassifyroadsigns,nonehaveadoptedanunsupervisedapproach.ThispaperproposesawayofSelf-Organizingfeaturemappingforrecognizingaroadsign.Theemergentself-organizingmap(ESOM)isemployedforthefeaturemappinginthisstudy.Ithasthecapabilityofvisualizingthedistancestructuresaswellasthedensitystructureofhigh-dimen-sionaldatasets,inwhichtheESOMissuitabletodetectnon-trivialclusterstructures.ThispaperdiscussestheusageofESOMforroadsigndetectionandclassification.Thebenchmarkingagainstsomeothercommonlyusedclassifierswasperformed.TheresultsdemonstratethattheESOMapproachoutperformstheothersinconductingthesamesimulationsoftheroadsignrecognition.WefurtherdemonstratethattheresultobtainedwithESOMissignif-icantlymoresuperiorthantraditionalSOMwhichdoesnottakeintotheboundaryeffectlikeESOMdid.KeywordsSelf-organizingmapDatavisualizationImageclassificationRoadsignrecognition1IntroductionSelf-organizingmap(SOM),proposedbyKohonen[1,2],canbeusedfordimensionreduction,vectorquantization,andvisualization.SomerecentSOM-basedapplicationsinimageprocessingandpatternrecognitiondomainscanbeseenin[3–5].SOMquantizesinputdatatoasmallnumberofneuronsandstillpreservesthetopologyofinputdata.Indeed,SOMcanbeseenasdiscreteapproximationofprincipalsurfacesininputspace[6,7].ManyvisualizationmethodsbasedonSOMwereproposed,forexamplesin[8–11].ConventionalSOMtopologyseemstobeinher-entlylimitedbythefixednetwork.Onemustadoptanumberoftrialsandteststoselectanappropriatenetworkstructureandsize.SeveralimprovedSOMsorrelatedalgorithmshavebeendevelopedtoovercometheseshort-comings.AllthesealgorithmsaremainlyinthedirectionofgrowinganSOMadaptively.Althoughmostoftheseextendedalgorithmsareabletodynamicallyincreasethenetworktoanappropriatesize,itmaynotbeeasytousethefinalSOMmapsforvisualizinghigh-dimensionalinputdataona2-Dplane,orfordistinguishingclustersona2-Dplane.Recently,theViSOM[12],anewvisualizationmethod,regularizestheinter-neurondistancessuchthattheinter-neurondistancesintheinputspaceresemblethoseintheoutputspaceafterthecompletionoftraining.Thisfeaturecanbeusefultosomeapplicationsbecauseitisabletopreservethetopologyinformationaswellastheinter-neurondistances.Thischaracteristicisattributedtotheoutputtopologypre-definedinaregular2-Dgridsothatthetrainedneuronsarealmostregularlydistributedintheinputspace.TheViSOMdeliversbetterdatavisualizationcomparedwithconventionalSOMandothervisualizationmethods.AnotherrecentapproachforSOM-basedvisualizationiscalledemergentSOM(ESOM)[13].EmergentSOMisanextensionofSOMthatallowstheemergenceofintrinsicstructuralfeaturesofhigh-dimensionaldataontoatwo-dimensionalmap.IthasbeendemonstratedthatusingY.-Y.NguwiS.-Y.Cho(&)DivisionofComputingSystems,SchoolofComputerEngineering,NanyangTechnologicalUniversity,NanyangAvenue,Singapore639798,Singaporee-mail:[e-mail]123NeuralComput&Applic(2010)19:601–615DOI10.1007/s00521-009-0315-6ESOMisasignificantlydifferentprocessfromusingk-means.ESOMisapowerfultoolforclustering,visualiza-tion,andclassification.Inthispaper,weproposethevisualizationofroadsignimagesthroughthemethodologyoffeatureclusteringandvisualizingbyemergentself-organisingmap(ESOM).Theobservationobtainedthroughthevisualizationprocesswillbediscussed.Sixclassesofroadsignsareinvestigated;theyarenamelystopsign,give-waysign,noleftturnsign,norightturnsign,speedlimit60km/handspeedlimit90km/h.ThefocusofthisworkistoshowhowtheESOMcanbeusedasanunsupervisednetworkthatisabletosegregatethesixclassesoftheroadsigns.Duetothelackofpubliclyavailableroadsigndatabase,thisworkdescribesandmakesavailablearoadsigndatabasethatconsistof447roadsceneimagesand1,600roadsignimages.Thepaperisorganizedasfollows:Sect.2describessomerelatedbackgroundworksandmotivationofpro-posingaroadsignrecognitionsystem.Section3givesanoverviewontheusageofESOM.Section4presentsthevisualizationandclassificationperformancesofusingESOMontheroadsignimage.Finally,conclusionofthispaperisdrawninSect.5.2RelatedworksandmotivationsInrecentyears,theworksondevelopingroadsignrec-ognitionsystemarenumbered.However,itisaveryimportantareathatdeserveswiderattention.Aroadsignrecognitionsystemprovidestimelyalerttowarnthedri-verofanycriticalsignahead.Theobjectiveofaroadsignrecognitionsystemistodetectandclassifyoneormoreroadsignsfromcolouredimagescapturedbycamera.Thereexistmanychallengesthatsuchasystemshouldaddress.Forinstances,lightingconditionisaverydifficultproblemtoregulate.Thestrengthofthelightdependsonthetimeofthedayandseason,andalsoontheweatherconditions.Inaddition,roadsignpatternswithinimagescanbeaffectedbyshadowsfromsur-roundingobjects.Ingeneral,aroadsignrecognitionsystemwillfirstdetecttheroadsignofinterestintheimagefollowedbyclassifyingitintodifferentclasses.Mostofthesolutionsrelyonth