搜索
OnTheCorrelationofImageSizetoSystemAccuracyinAutomaticFingerprintIdentificationSystemsJ.K.Schneider ,C.E.Richardson,F.W.Kiefer,andVenuGovindaraju Ultra-ScanCorporation,4240RidgeLeaRd,Amherst,NewYork14226 CenterofExcellenceforDocumentAnalysisandRecognition(CEDAR)StateUniversityofNewYorkatBuffalo,Buffalo,NY14228,U.S.A.zshi@cedar.buffalo.edu,(AFIS)isdependentonmanyvariablesrangingfromthequalityofthefrictionskinsur-faceitself,totheabilityoftheimageacquisitiondevicetoaccuratelyimagethefingerprint,tothefeatureextractandmatchsoftwaretodetectthefeaturesetanddeclareacorrelation(amatch/nomatch)basedonsomecriteria.Thegene-sisofthesesystemsexclusivelyusedrolledimpressionfingerprintimagery.AsAFISdevelopersattemptedtoreducethecostandimprovetheusabilityofthesesystemsforcommercialapplications,plainimpressionimagerywasadoptedandtotalscannedimageareawassignificantlyreduced.Thispaperwillattempttoquantifytheeffectsofreducedimageareaonoverallsystemaccuracyinasinglefingerplainimpressionenvironment.TheresultscanbedirectlyappliedbyAFISdesignersastheyattempttomeetsystemsrequirementswithrespecttokeyaccu-racyparameterssuchasfalsematchrates,falsenon-matchratesandtheresultingworkloadcausedbythesystem.1IntroductionAutomaticFingerprintIdentificationSystemsdevelopedforlawenforcementapplica-tionsstrictlyusedrolledimpressionfingerprintimagery,capturingasmuchfrictionskinsurfaceinformationthatisavailableontheparticularfinger,andimagedallavailablefingers.Therequirementsforthelive-scanimageacquisitiondevicearewelldefinedandpublishedaspartofaNISTstandardreferredtoasAppendixF.[1]Scanningthismuchimageareaensuredthecaptureofthecoreanddeltaofthefingerprintwhichen-ablesclassificationoftheimage.ClassificationallowstheAFIStoreducethesizeofthesearchspacethroughpenetrationintothedatabaseinaone-to-manymatchingconfig-uration,bysegmentingthedatabaseaccordingtoclassification.Theexpectedaccuracyofsuchsystemhasbeenwellformulatedasdiscussedin[2]andgivenas: (1) !(Equation(1)onPage2)2where is,forafixedcomparisonscorethreshold,theprobabilityofafalsenon-matchwhencomparinganytwoormoremathchingtemplatesinacoldsearch,i.e.theprobabilitythatwedonotmatchwhenweshould. is,forafixedcomparisonscorethreshold,theprobabilityofafalsematchwhencomparinganytwoormorenon-mathchingtemplatesinacoldsearch,i.e.theprobabilitythatwedomatchwhenweshouldnot. isthenumberoftemplatesinthedatabase. istheprobabilityofselectingthewrongclassificationbin.Whileclassificationofthefingerprintreducesoverallsearchtimes,thisapproachhasanimpactonoverallsystemaccuracy,specificallyincreasingfalsenon-matchrates,dependingontheaccuracyoftheclassificationschemeasdiscussedin[3]andgivenas: (2) ! (Equation(2)onPage2)where isthenumberoftemplatesintheselectedclassificationbin.Thelive-scanrolledimpressionfingerprintscannersusedforcapturinganimageofthefingerrequiretheusertorollthefingerfromsidetoside,generallysynchro-nizingthespeedoftherollwithsomeguidinglightsorLEDs.Thisprocesscanbequitedemandingontheactualuserandgenerallyrequiresatrainedattendanttoassistintheoverallprocess.Thesefactorsareoftendeemedtoocomplexortoocostlyforcommercialapplications.Withthegrowthinpopularityofthesesystemsforavarietyofnon-lawenforcementapplications,developersconcentratedonimprovingtheusabil-ityofthesesystemsthroughcapturingplainimpressionimageryasopposedtorolledimpressionimagery.Thisreducedthecostandcomplexityoftheimageacquisitionde-viceconsiderably,whileimprovingtheireaseofuse.Aseffortsweremadetofurtherreducecosts,thetotalfrictionskinsurfaceareaimagedbytheplainimpressionfinger-printreadercontinuedtodecrease.Thetotalareacapturedvariesfromvendortovendorsincetherearevirtuallynostandardsgoverningtheattributesofplainimpressionfin-gerprintreaders.Recently,initiativesbyAAMVA,theAmericanAssociationofMotorVehiclesAdministrationhaveintroducedsuchcriteriabutthesestandardshavenotbeenadopteduniversallybyallcommercialapplications.[3]MostAFISdeployedforcommercialapplications,usingplainimpressionimagery,donotattempttoclassifythefingerprintimageduetothelackofreliablyknowingthelocationofthedelta.Thiseliminatestheerrorrateduetopenetrationofthedatabases.However,theeffectsonverifyingreducedimageareahavenotbeenfullycharacterizedfortheresultingsystemsrequiringaone-to-Nsearch.Quiteoften,seeminglyintuitivestatementswillbemadethatthereducedimageareaisgoodenoughforagivenappli-cationduetothereducedsizeofthedatabase.Whileitistruethateachapplicationshasdifferentrequirementswithrespecttofalsematchandfalsenon-matchperformance,un-derstandingbestcaseanticipatedperformanceisessentialpriortodesigningtheoverallsystem.32PerformanceModel3TestDataPlainimpressionfingerprintimageswerecollectedusingUltra-ScansModel503Ul-trasonicFingerprintReader.Thisdevicecapturesalarge,plainimpressionimageareaof2.03cm3.05cm(0.8x1.2)ataresolutionof1,270samples/c