A review on Gabor wavelets for face recognition

AbstractDuetotherobustnessofGaborfeaturesagainstlocaldistortionscausedbyvarianceofillumi-nation,expressionandpose,theyhavebeensuccess-fullyappliedforfacerecognition.TheFacialRecognitionTechnology(FERET)evaluationandtherecentFaceVerificationCompetition(FVC2004)haveseenthetopperformanceofGaborfeaturebasedmethods.Thispaperaimstogiveadetailedsurveyofstateoftheart2DfacerecognitionalgorithmsusingGaborwaveletsforfeatureextraction.Existingprob-lemsarecoveredandpossiblesolutionsaresuggested.KeywordsJointtime–frequencyanalysisÆGaborwaveletsÆFacerecognition1IntroductionAutomaticrecognitionofhumanfaceshasbeenanactiveresearchareainrecentyears.Inadditiontotheimportanceofpureresearch,ithasanumberofcom-mercialandlaw-enforcementapplicationssuchassurveillance,security,telecommunicationsandhuman–computerintelligentinteractionetc.Variousap-proachesforfacerecognitionhavebeenproposedandtheycanberoughlyclassifiedintoeitheranalyticapproachesorholisticapproaches.Analyticapproachesusethingssuchasdistancesandanglesbetweenfiducialpointsontheface,shapesoffacialfeaturesandlocalfeatures.Forexampleintensityvaluesextractedfromfacialfeaturescanbeused.Themainadvantageofanalyticapproachesistoallowforaflexibledeformationatthekeyfeaturepointssothatposechangescanbecompensatedfor.In[1]facialregionsarematchedwithtemplatesofeyes,nosesandmouthsrespectivelyandfacerecognitionisperformedwithoutgeometricalconstraints.AsetofSVMclassi-fiersisappliedtoextractdifferentfacialcomponentsandthegrayvaluesofeachcomponentarecombinedintoasinglefeaturevector[2].Theembedded2DHiddenMarkovModelisadoptedin[3],inwhichanoverlappedwindowisshiftedoverthefaceimageandtheDCTcoefficientsarecomputedandfedintothemodelastheobservationvector.Whileanalyticapproachescomparethesalientfacialfeaturesdetectedfromtheface,holisticapproachesmakeuseoftheinformationderivedfromthewholeface.PrincipalComponentsAnalysis(PCA)isatypicalholisticmethod,whichisastatisticaltechniqueusingtheKarhunen–Loevetransform.TurkandPentland[4]developedawellknowneigenfacemethodforbothfacerepresentationandrecognitionusingthePCAtechnique.PCAcanachievetheoptimalrepresenta-tioninthesenseofmean-squareerror,butthediffer-encesbetweenfacesfromdifferentpeopleseemmoreimportantinfacerecognition[5].Basedonthisobservation,LinearDiscriminantAnalysis(LDA)[6]isappliedfortheFisherface[7]methods.LDAdefinesaprojectionthatmakesthewithin-classscattersmallandthebetween-classscatterlarge,butitrequireslargetrainingsamplesetsforgoodgeneralization,whichareusuallynotavailableforfacerecognitionL.Shen(&)ÆL.BaiSchoolofComputerScienceandInformationTechnology,UniversityofNottingham,Nottingham,UKe-mail:[e-mail]-mail:[e-mail](2006)9:273–292DOI10.1007/s10044-006-0033-y123SURVEYAreviewonGaborwaveletsforfacerecognitionLinlinShenÆLiBaiReceived:6December2004/Accepted:17May2006/Publishedonline:18August2006
Springer-VerlagLondonLimited2006applications.Asaresult,PCAisnormallyadoptedtoreducethefeaturedimensionbeforeLDAcanbeap-plied[8].OthertechniquesproposedinliteraturetosolvetheSmallSampleSize(SSS)problemareRegu-larizedLDA(RLDA)[9],EnhancedLDA(ELDA)[10]andDirectLDA(DLDA)[11].Neuralnetworks[12–14]havealsobeenusedtoclassifyglobalfeatures.Globaltechniquesworkwellforfrontalviewfaceimages,buttheyaresensitivetotranslationandrota-tionetc.ofthepose[2].Usuallynormalizationisanimportantandinevitableprocessforthesemethods,wherebyasmallnumberofprominentpointsinthefacesuchastheeyes,nostrilsorcenterofthemouthareusedtoresizeandrotatetheinputfaceimage.Afternormalization,theinputfaceimagecanbealignedwiththemodelfaceandthenrecognitioncanbeperformed.Moredetailedliteratureonfacerecog-nitionapproachescanbefoundin[15–17].Despiteremarkableprogressessofar,thegeneraltaskoffacerecognitionremainsachallengingproblem,thisismainlyduetothecomplexdistortionsthatcanbecausedbyvariationsinillumination,facialexpres-sionsandposes.Itiswidelybelievedthatlocalfeaturesinfaceimagesaremorerobustagainstsuchdistortionsandaspatial–frequencyanalysisisoftendesirabletoextractsuchfeatures[16,18].Withgoodcharacteristicsofspace–frequencylocalization,waveletanalysisistherightchoiceforthispurpose[19,20].Inparticular,amongvariouswaveletbasesGaborfunctionsprovidetheoptimizedresolutioninboththespatialandfre-quencydomains[21,22].Gaborwaveletsseemtobetheoptimalbasistoextractlocalfeaturesforpatternrecognition,forseveralreasons:•Biologicalmotivation:theshapesofGaborwave-letsaresimilartothereceptivefieldsofsimplecellsintheprimaryvisualcortex[22].•Mathematicalmotivation:theGaborwaveletsareoptimalformeasuringlocalspatialfrequencies[23,24].•Empiricalmotivation:Gaborwaveletshavebeenfoundtoyielddistortiontolerantfeaturespacesforotherpatternrecognitiontasks,includingtexturesegmentation[25,26],handwrittennumeralrecog-nition[27]andfingerprintrecognition[28].TheapplicationofGaborwaveletsforfacerecog-nitionhasbeenpioneeredbyLadesetal.’sworksinceDynamicLinkArchitecture(DLA)wasproposedin1993[29].Inthissystem,facesarerepresentedbyarectangulargraphwithlocalfeaturesextractedatdeformablenodesusingGabo