AppearedinProc.ofInternationalConferenceonPatternRecognition(ICPR),Que - PDF document

AppearedinProc.ofInternationalConferenceonPatternRecognition(ICPR),QuebecCity,August11-15,2002.AHybridFingerprintMatcherArunRoss,AnilJainMichiganStateUniversityEastLansing,MI,USA48824frossarun,jaing@cse.msu.eduJamesReismanSiemensCorporateResearch,Inc.Princeton,NJ,USA08540james.reisman@scr.siemens.comAbstractWedescribeahybridngerprintmatchingschemethatusesbothminutiaeandridgeowinformationtorepresentandmatchngerprints.Asetof8Gaborlters,whosespa-tialfrequenciescorrespondtotheaverageinter-ridgespac-inginngerprints,isusedtocapturetheridgestrengthatequallyspacedorientations.Asquaretessellationofthel-teredimagesisthenusedtoconstructaneight-dimensionalfeaturemap,calledtheridgefeaturemap.Theridgefea-turemapalongwiththeminutiaesetofangerprintimageisusedformatchingpurposes.Thegenuineacceptrateofthehybridmatcherisobservedtobe10%higherthanthatofaminutiae-basedmatcheratlowfalseacceptrates.FingerprintvericationusingthehybridmatcheronaPen-tiumIII(800MHz)processor,takes1:4seconds.1.IntroductionFingerprintmatchingtechniquescanbebroadlyclas-siedasbeingminutiae-basedorcorrelation-based.Minutiae-basedtechniquesattempttoaligntwosetsofminutiaepointsanddeterminethetotalnumberofmatchedminutiae[6].Correlation-basedtechniques,ontheotherhand,comparetheglobalpatternofridgesandfurrowstoseeiftheridgesinthetwongerprintsalign[1].Theper-formanceofminutiae-basedtechniquesreliesontheaccu-ratedetectionofminutiaepointsandtheuseofsophisti-catedmatchingtechniquestocomparetwominutiaeeldswhichundergonon-rigidtransformations.Theperformanceofcorrelation-basedtechniquesisaffectedbynon-lineardistortionsandnoisepresentintheimage.Ingeneral,ithasbeenobservedthatminutiae-basedtechniquesperformbetterthancorrelation-basedones.Jainetal.[5]haveproposedanovelrepresentationschemethatcapturesglobalandlocalfeaturesofanger-printinacompactxedlengthfeaturevector,calledtheFingerCode.Thistechniqueviewsangerprintasanori-entedtexture(Figure1)andtheirgenericrepresentationofX(a)(b)(c)(d)Figure1.Fingerprintasanorientedtexturepattern.(a)Angerprintwiththecore()andfourminutiaepoints(ridgeending­;ridgebifurcation­)markedonit;(b)theconstantinter­ridgespacinginalocalregionofthengerprint;(c)thedominantdirectionoftheridgesin(b);(d)thepowerspectrumof(b).orientedtexturereliesonextractingacorepointinthen-gerprint.Theirtechnique,however,suffersfromthefollow-ingshortcomings:(i)Theframeofreferenceisbasedonaglobalsingularpointi.e.,thecorepoint(Figure2(a)).De-tectionofthecorepointisnon-trivial;thecorepointmaynotevenbepresentinsmall-sizedimagesobtainedusingsolid-statesensors.(ii)Thengerprintalignmentisbasedonasinglereferencepointandis,therefore,notveryro-bustwithrespecttoerrorsinthelocationofthereferencepoint.(iii)Thetessellationdoesnotcovertheentireim-age.Furthermore,ifthecoreweretobedetectedcloseto AppearedinProc.ofInternationalConferenceonPatternRecognition(ICPR),QuebecCity,August11-15,2002.theboundaryoftheimage,thetessellationwillincludeanextremelysmallportionoftheimage.X(a)(b)Figure2.Tessellatingthengerprintimage.(a)Circulartessellation(80sectors)aboutacorepoint.(b)Squaretessellation(169cells)overtheentireimage.Wepresentangerprintrepresentationschemethatcon-structsaridgefeaturemapbyobservingthelocalridgeori-entation.ThelocalridgecharacteristicsareextractedviaasetofGaborlterswhosefrequencycorrespondstotheinter-ridgespacinginngerprints.Unlikein[5],thel-teringisdoneontheenhancedimagesratherthantherawinputimages.Insteadofusingcirculartessellation,asquaretessellationisused(Figure2(b)).Thetessellationincludestheentireimage,andallthetessellatedcellsareofthesamesize.Thetessellationisnotbasedondetectinganyland-markpoints.Thengerprintimagesarealignedusingtheoverallminutiaeinformation;thisismorerobustthanusingonlythecorepointforaligningimagepairsasdonein[5].2.RidgeFeatureMapsBytuningaGaborltertoaspecicfrequencyanddi-rection,textureinformationfromimagescanbeextracted[2].AnevensymmetricGaborlterhasthefollowinggen-eralforminthespatialdomain:G;f(x;y)=exp12x022x+y022ycos(2fx0);x0=xsin+ycos;y0=xcosysin;wherefisthefrequencyofthesinusoidalplanewaveatananglewiththex-axis,andxandyarethestandarddeviationsoftheGaussianenvelopealongthexandyaxes,respectively.Forthe240240imagesobtainedusingtheVeridicomsensor(seesection4),theGaborparametersaresettothefollowingvalues:(i)f=0:125(correspondstoaninter-ridgedistanceof8);(ii)x=y==4(basedonempiricaldata);(iii)=f0;22:5;45;67:5;90;112:5;135;157:5g,re-sultingineightGaborlters.(a)(b)Figure3.Fingerprintimage(a)beforeand(b)afterenhancement.Weenhancetheinputimage[3]priortoltering(inordertoimprovetheclarityofridgesandfurrows).Theminutiaefeaturesarealsoextractedafterprocessingtheenhancedn-gerprintimage.Figure3showsangerprintimagebeforeandafterenhancement.Filteringrequiresconvolvingtheenhancedimagewitheachofthe8Gaborltersinthespa-tialdomain.Inordertospeed-uptheltering,theconvo-lutionisperformedinthefrequencydomain.Eightlteredimagesareobtainedasaresultofthisltering(Figure4).(a)(b)(c)(d)Figure4.FilteringoftheimageinFigure3(b).Only4ofthe8lteredimagesareshown.Whilealteredimageinitsentiretycanbeusedasarep-resentationscheme,thepresenceoflocaldistortionswouldaffectthematchingprocessdrastically.Moreover,itisthelocalvariationsinridgestructure(combinedwiththeglobalridgeconguration)thatwillprovideabetterrepresentationofthengerprint.Toexaminelocalvariations,theimageistessellatedintosquarecells,andfeaturesfromeachofthecellsarecomputed.Thesizeofacellischosentocorre-spondtotwicetheinter-ridgespacing(1616).A16-pixelwideborderoftheimageisnotincludedinthetessellation.Thisresultsinnc=13cellsineachrowandcolumnofthesquaregrid,withatotalof169cells.Thevarianceofthe AppearedinProc.ofInternationalConferenceonPatternRecognition(ICPR),QuebecCity,August11-15,2002.pixelintensitiesineachcellofall8lteredimagesisusedasafeaturevector.Thevariancecorrespondstotheenergyofthelterresponse,andis,therefore,ausefulmeasureofridgeorientationinalocalneighborhood.Thosetessellatedcellsthatcontainacertainproportionofbackgroundpixelsarelabeledasbackgroundcellsandthecorrespondingfea-turevalueissettozero.Aneight-dimensionalfeaturemap(1313image)correspondingtothe8lteredimagesisob-tainedinthisway(Figure5).Theseridgefeaturemapsareusedtorepresentandmatchaqueryimagewithatemplate.(a)(b)(c)(d)Figure5.Featuremaps(1313images)forthe4lteredimagesingure4.3.HybridFingerprintMatcherThehybridngerprintmatcherproposedhereutilizestwodistinctsetsofngerprintinformation:minutiaefea-tures,andridgefeaturemaps.Minutiaeinformationisob-tainedusingthetechniquedescribedin[4].Whenaqueryimageispresented,thematchingproceedsasfollows:(i)thequeryandtemplateminutiaefeaturesarematchedtogenerateaminutiaematchingscoreandanafnetransfor-mationthatalignsthequeryandtemplatengerprints;(ii)thequeryimageislteredusing8Gaborlters;(iii)theridgefeaturemapisextractedfromtheselteredimages;(iv)thequeryandtemplateridgefeaturemapsarematched;(v)theminutiaeandridgefeaturemapmatchingscoresarecombinedtogenerateasinglematchingscore(Figure6).Forcomparingtheridgefeaturemapsoftwoimages,itisnecessarythattheimagesthemselvesarealignedappro-priatelytoensureanoverlapofcommonregioninthetwongerprintimages.Thisisdonebydeterminingtheafnetransformationparameters,(tx;ty;t),thatwouldalignthequeryimagewiththetemplate.Thetransformationparam-etersarecomputedbymatchingthetwosetsofminutiaepointsusinganelasticpointmatchingalgorithm[4].In-steadofrotatingthequeryimagebytandthenlteringit,theGaborltersareappropriatelyrotatedandthemodiedGaborlterbankisappliedtothequeryimage.Theresult-inglteredimagesarethenrotatedbytandtranslatedby(tx;ty).Thisreducestheartifactsassociatedwithlteringarotatedimage.Theminutiaematchingscoreisameasureofthesimilar-ityoftheminutiaesetsofthequeryandtemplateimages.Thesimilarityscoreisnormalizedinthe[0;100]range.Theridgefeaturemapsofthequeryandthetemplateim-agesarecomparedbycomputingthesumoftheEuclideandistancesofthe8-dimensionalfeaturevectorsinthecorre-spondingtessellatedcells.(Cellsthataremarkedasback-ground,arenotusedinthematchingprocess).Thedistancescoreisnormalizedinthe[0;100]range,andconvertedtoasimilarityscorebysimplysubtractingitfrom100.LetSMandSRindicatethesimilarityscoresobtainedusingminutiaematchingandridgefeaturemapmatching,respec-tively.Then,thenalmatchingscore,S,iscomputedas,S=SM+(1)SR,where2[0;1].Fortheexperi-mentalresultsreportedinthispaper,wassetto0:5.4.ExperimentsandResultsThengerprintdatabaseusedinourexperimentscon-sistsofngerprintimpressionsobtainedfrom160non-habituated,cooperativesubjectsusingtheVeridicomsensor(300300imagesat500dpi).Thedatawascollectedovertwosessions.Ineachofthesessions,asubjectwasaskedtoprovide2impressionsofeachof4differentngers-theleftindexnger,theleftmiddlenger,therightindexngerandtherightmiddlenger.Atotalof2;560imageswereacquiredovertwotimesessions.Thisisadifcultdatabaseforangerprintmatcherduetothefollowingreasons:(i)Thesmall-sizedsensorcapturesonlyalimitedportionofthesubject'sngerprint,andmultipleimpressionsofthesamenger,insomecases,hadverylittleoverlap.(ii)Manysubjectswereobservedtohavedryngersthatresultedinpartialorfaintprints.The300300imageswererstre-sizedto240240(inter-ridgedistancewasreducedfrom10pixelsto8pixels),andthenpaddedwithzerostoexpandthemto256256,inordertospeed-uptheFourieropera-tions.TheROC(ReceiverOperatingCharacteristic)curvesdepictingtheperformancesoftheminutiae,ridgefeaturemapandhybridmatchersareshowninFigure7.Thehy-bridtechniqueoutperformstheminutiae-basedschemeoverawiderangeofFARvalues.Theequalerrorrateofthehy-bridtechniqueisobservedtobe4:5%,whilethatoftheminutiae-basedmatcheris14%.Theexperimentsalsoshowthattheminutiaeinformationandridgeowinforma-tioncomplementeachother.TheexperimentsreportedherewereconductedonaPen- AppearedinProc.ofInternationalConferenceonPatternRecognition(ICPR),QuebecCity,August11-15,2002.Query imageScoreScoreMatchingMinutiae setRidge feature mapMinutiae setRidge feature mapTemplateQueryScoreMatchingSum RuleMinutiaeFeature MapMatchingTransformation parameters to align query with templatebefore extracting ridge feature map of queryFigure6.Thematchingprocess.10-11001016065707580859095100False Accept Rate (%)Genuine Accept Rate (%)HybridRidge Feature MapMinutiae Equal Error LineFigure7.ROCshowingtheperformancesofthethreematchers.tiumIII,800Mhzprocessor,runningWindows2000.Thetotaltimeforngerprintverication(one-to-onematching)was1:4seconds.5.SummaryandFutureWorkWehaveproposedangerprintmatchingtechniquethatcombinesminutiaeinformationwiththeridgeowinfor-mation.Experimentsindicatethatthehybridtechniqueper-formsmuchbetterthanapurelyminutiae-basedmatchingscheme.Currently,theminutiaeinformationisusedtoalignthequeryandthetemplateimages,beforecomputingtheridgefeaturemap.Weareworkingonnon-minutiaebasedalignmenttechniquesthatmakeuseoforientationeldandridgemapinformationtoalignimagepairs.References[1]A.M.Bazen,G.T.B.Verwaaijen,S.H.Gerez,L.P.J.Vee-lenturf,andB.J.vanderZwaag.Acorrelation-basednger-printvericationsystem.InProceedingsoftheProRISC2000WorkshoponCircuits,SystemsandSignalProcessing,Veld-hoven,Netherlands,Nov2000.[2]J.Daugman.Recognizingpersonsbytheiririspatterns.InA.K.Jain,R.Bolle,andS.Pankanti,editors,Biometrics:PersonalIdenticationinaNetworkedSociety,pages103–121.KluwerAcademicPublishers,1999.[3]L.Hong,Y.Wan,andA.K.Jain.Fingerprintimageenhance-ment:Algorithmsandperformanceevaluation.IEEETrans-actionsonPAMI,20(8):777–789,Aug1998.[4]A.K.Jain,L.Hong,andR.Bolle.On-linengerprintveri-cation.IEEETransactionsonPAMI,19(4):302–314,April1997.[5]A.K.Jain,S.Prabhakar,L.Hong,andS.Pankanti.Filterbank-basedngerprintmatching.IEEETransactionsonImageProcessing,9(5):846–859,May2000.[6]L.O'Gorman.Fingerprintverication.InA.K.Jain,R.Bolle,andS.Pankanti,editors,Biometrics:PersonalIden-ticationinaNetworkedSociety,pages43–64.KluwerAca-demicPublishers,1999.