Local Gabor Binary Pattern Histogram Sequence (LGB

July29,200722:47WSPC/164-IJIG00291InternationalJournalofImageandGraphicsVol.7,No.4(2007)777–793c
WorldScientificPublishingCompanyLOCALGABORBINARYPATTERNSBASEDONMUTUALINFORMATIONFORFACERECOGNITIONWENCHAOZHANGSchoolofComputerScienceandTechnologyHarbinInstituteofTechnologyHarbin,150006,[e-mail]∗,XILINCHEN†andWENGAO‡ICT-ISVISIONFRJDL,InstituteofComputingTechnologyChineseAcademyofScience,Beijing,100080,P.R.China∗[e-mail]†[e-mail]‡[e-mail]11November2006Revised18March2007Accepted19March2007Appropriaterepresentationisoneofthekeystothesuccessoffacerecognitiontech-nologies.Inthispaper,wepresentanovelfacerepresentationapproachusingareducedsetoflocalhistogramsbasedonLocalGaborBinaryPatterns(LGBP).Inthepro-posedmethod,afaceimageisfirstrepresentedbytheLGBPhistogramswhichareextractedfromtheLGBPimages.Then,thelocalLGBPhistogramswithhighsepa-rabilityandlowrelevanceareselectedtoobtainadimension-reducedfacedescriptor.Extensiveexperimentalresultsdemonstratethattheproposedmethodnotonlygreatlyreducesthedimensionalityoffacerepresentation,butalsooutperformsthestate-of-the-artapproachesforfacerecognition,suchasFisherfaces,andGaborFisherClassification(GFC).Keywords:Localbinarypatterns(LBP);Gaborwavelets;mutualinformation;Fisherlineardiscriminant;localGaborbinarypatterns(LGBP).1.IntroductionFacerecognitionhasbeenoneofthemostchallengingandactiveresearchtopicsincomputervisionforseveraldecadesduetoitsscientificvaluesandwidepotentialapplications.Muchprogresshasbeenmadeinthelastdecade.1,2However,thegen-eralproblemoffacerecognitionremainsunsolved,sincemostofthesystemstodatecanonlysuccessfullyrecognizefaceswhenimagesareobtainedunderconstrainedconditions.Theirperformancewilldegradeabruptlywhenfaceimagesarecapturedundervaryinglightingconditions,poses,expressions,agesandsoon.777July29,200722:47WSPC/164-IJIG00291778W.Zhangetal.Inearlierworks,geometricfeature-basedmethods3–7havebeenwidelyinves-tigated.Feature-basedmethodsusepropertiesandrelations(e.g.distancesandangles)betweenfacialfeatures,suchaseyes,mouth,nose,andchintoperformrecognition.Oneofthemostsuccessfulsystemsiselasticbunchgraphmatching(EBGM)system,6whichisrobusttoilluminationchange,translation,distortion,rotation,andscaling.Althoughthefeature-basedrepresentationmethodsareinsen-sitivetovariationsinilluminationandpose,precisealignmentandfacialfeatureextractionprocess,however,arecriticalfortheirperformance.Later,appearance-basedmethodshavebeenintroducedwhichuselowdimen-sionalrepresentationsofobjectstoperformrecognition.8–15Eigenfaces8andFisherfaces12havedemonstratedthepowerofappearance-basedmethodsbothineaseofimplementationandinrecognitionaccuracy.Theirperformance,however,willbedegradedwhenthedistributionofthetestimagesisdifferentfromthatofthetrainingimages.Oneofthemaindifficultiesforfacerecognitionarisesfromlargewithin-classvariations(duetoillumination,facialexpression,aging)andrathersmallbetween-classvariations(duetosimilarityofindividualappearances)inhumanfaceimages.Thesevariationsincludethelocalvariations(e.g.wrinklesappearingatthemouthcorner)andtheglobalvariations(e.g.lightingcanchangethewholevariationoffaceimage).Therefore,robustfacerepresentationagainstfacialvariationsiscriticalforapracticalfacerecognitionsystem.Recently,localbinarypatterns(LBP)operatorhasbeensuccessfullyusedforfacedetection16andrecognition.17FacerepresentationwithLBPencodesboththelocalandglobalinformationbyaconcatenatedLBPhistogram.FacialfeatureextractedbytheLBPoperatorisrobusttoilluminationvariationsbecausetheLBPfeaturesareinvarianttothemonotonicgray-scalechanges.Theauthorsreportedthestate-of-the-artresultsontheFERETfacedatabase.However,undertheconditionofvaryinglightingandaging,itsperformanceisstillnotsatisfactory.Sincemultiresolutionhistogramscouldimprovetheperformanceofobjectclassification,18meanwhile,Gaborbasedfacerepresentationisrobusttoillumina-tionvariations6andefficienttodescribelocalimagefeatures.19WehaveproposedcombiningGaborwaveletswithLBPoperatortorepresentfaceimagetoobtainrobustfeatureagainstfacialvariations.ThecombiningoperatoristermedaslocalGaborbinarypatterns(LGBP)20operator.Inthismethod,firstly,weobtainthemultiresolutionimagesbyconvolvingthefaceimagewithmulti-scaleandmulti-orientationGaborfilters.Secondly,LBPoperatorisconductedonthemultireso-lutionimagestoobtaintheLGBPimages.Thirdly,localhistogramsareextractedfromtheLGBPimagesandallthelocalLGBPhistogramsareconcatenatedintoonehistogramtorepresentthegivenfaceimage.Experimentalresultshavedemon-stratedthattheperformanceoffacerecognitionwithLGBPissuperiortoboththeLBP-basedapproach17andGabor-basedapproach.19However,facerepresentationwithLGBPishighdimensionalduetothemultipleGabortransformationsofLGBPoperator.ThusfurtherdimensionalityreductionJuly29,200722:47WSPC/164-IJIG00291LocalGaborBinaryPatternsforFaceRecognition779isnecessaryafterobtainingtheLGBPhistograms.Therearetwomajorcategoriesofmethodsofdimensionalityreduction,featureselectionandfeaturetransform.Featureselectionmethodskeeponlyusefulfeaturesanddis