人脸识别文献翻译中英双文.docx

人脸识别文献翻译中英双文.docx

《人脸识别文献翻译中英双文.docx》由会员分享，可在线阅读，更多相关《人脸识别文献翻译中英双文.docx（12页珍藏版）》请在冰豆网上搜索。

人脸识别文献翻译中英双文

4Two-dimensionalFaceRecognition

4.1FeatureLocalization

Beforediscussingthemethodsofcomparingtwofacialimageswenowtakeabrieflookatsomeatthepreliminaryprocessesoffacialfeaturealignment.Thisprocesstypicallyconsistsoftwostages:

facedetectionandeyelocalization.Dependingontheapplication,ifthepositionofthefacewithintheimageisknownbeforehand（foracooperativesubjectinadooraccesssystemforexample）thenthefacedetectionstagecanoftenbeskipped,astheregionofinterestisalreadyknown.Therefore,wediscusseyelocalizationhere,withabriefdiscussionoffacedetectionintheliteraturereview.

Theeyelocalizationmethodisusedtoalignthe2Dfaceimagesofthevarioustestsetsusedthroughoutthissection.However,toensurethatallresultspresentedarerepresentativeofthefacerecognitionaccuracyandnotaproductoftheperformanceoftheeyelocalizationroutine,allimagealignmentsaremanuallycheckedandanyerrorscorrected,priortotestingandevaluation.

Wedetectthepositionoftheeyeswithinanimageusingasimpletemplatebasedmethod.Atrainingsetofmanuallypre-alignedimagesoffacesistaken,andeachimagecroppedtoanareaaroundbotheyes.Theaverageimageiscalculatedandusedasatemplate.

Figure4-1Theaverageeyes.Usedasatemplateforeyedetection.

Botheyesareincludedinasingletemplate,ratherthanindividuallysearchingforeacheyeinturn,asthecharacteristicsymmetryoftheeyeseithersideofthenose,provideausefulfeaturethathelpsdistinguishbetweentheeyesandotherfalsepositivesthatmaybepickedupinthebackground.Althoughthismethodishighlysusceptibletoscale（i.e.subjectdistancefromthecamera）andalsointroducestheassumptionthateyesintheimageappearnearhorizontal.Somepreliminaryexperimentationalsorevealsthatitisadvantageoustoincludetheareaofskinjustbeneaththeeyes.Thereasonbeingthatinsomecasestheeyebrowscancloselymatchthetemplate,particularlyifthereareshadowsintheeye-sockets,buttheareaofskinbelowtheeyeshelpstodistinguishtheeyesfromeyebrows（theareajustbelowtheeyebrowscontaineyes,whereastheareabelowtheeyescontainsonlyplainskin）.

Awindowispassedoverthetestimagesandtheabsolutedifferencetakentothatoftheaverageeyeimageshownabove.Theareaoftheimagewiththelowestdifferenceistakenastheregionofinterestcontainingtheeyes.Applyingthesameprocedureusingasmallertemplateoftheindividualleftandrighteyesthenrefineseacheyeposition.

Thisbasictemplate-basedmethodofeyelocalization,althoughprovidingfairlypreciselocalizations,oftenfailstolocatetheeyescompletely.However,weareabletoimproveperformancebyincludingaweightingscheme.

Eyelocalizationisperformedonthesetoftrainingimages,whichisthenseparatedintotwosets:

thoseinwhicheyedetectionwassuccessful;andthoseinwhicheyedetectionfailed.Takingthesetofsuccessfullocalizationswecomputetheaveragedistancefromtheeyetemplate（Figure4-2top）.Notethattheimageisquitedark,indicatingthatthedetectedeyescorrelatecloselytotheeyetemplate,aswewouldexpect.However,brightpointsdooccurnearthewhitesoftheeye,suggestingthatthisareaisofteninconsistent,varyinggreatlyfromtheaverageeyetemplate.

Figure4-2–Distancetotheeyetemplateforsuccessfuldetections（top）indicatingvarianceduetonoiseandfaileddetections（bottom）showingcrediblevarianceduetomiss-detectedfeatures.

Inthelowerimage（Figure4-2bottom）,wehavetakenthesetoffailedlocalizations（imagesoftheforehead,nose,cheeks,backgroundetc.falselydetectedbythelocalizationroutine）andonceagaincomputedtheaveragedistancefromtheeyetemplate.Thebrightpupilssurroundedbydarkerareasindicatethatafailedmatchisoftenduetothehighcorrelationofthenoseandcheekboneregionsoverwhelmingthepoorlycorrelatedpupils.Wantingtoemphasizethedifferenceofthepupilregionsforthesefailedmatchesandminimizethevarianceofthewhitesoftheeyesforsuccessfulmatches,wedividethelowerimagevaluesbytheupperimagetoproduceaweightsvectorasshowninFigure4-3.Whenappliedtothedifferenceimagebeforesummingatotalerror,thisweightingschemeprovidesamuchimproveddetectionrate.

Figure4-3-Eyetemplateweightsusedtogivehigherprioritytothosepixelsthatbestrepresenttheeyes.

4.2TheDirectCorrelationApproach

Webeginourinvestigationintofacerecognitionwithperhapsthesimplestapproach,knownasthedirectcorrelationmethod（alsoreferredtoastemplatematchingbyBrunelliandPoggio）involvingthedirectcomparisonofpixelintensityvaluestakenfromfacialimages.Weusetheterm‘DirectCorrelation’toencompassalltechniquesinwhichfaceimagesarecompareddirectly,withoutanyformofimagespaceanalysis,weightingschemesorfeatureextraction,regardlessofthedistancemetricused.Therefore,wedonotinferthatPearson’scorrelationisappliedasthesimilarityfunction（althoughsuchanapproachwouldobviouslycomeunderourdefinitionofdirectcorrelation）.WetypicallyusetheEuclideandistanceasourmetricintheseinvestigations（inverselyrelatedtoPearson’scorrelationandcanbeconsideredasascaleandtranslationsensitiveformofimagecorrelation）,asthispersistswiththecontrastmadebetweenimagespaceandsubspaceapproachesinlatersections.

Firstly,allfacialimagesmustbealignedsuchthattheeyecentersarelocatedattwospecifiedpixelcoordinatesandtheimagecroppedtoremoveanybackgroundinformation.Theseimagesarestoredasgrayscalebitmapsof65by82pixelsandpriortorecognitionconvertedintoavectorof5330elements（eachelementcontainingthecorrespondingpixelintensityvalue）.Eachcorrespondingvectorcanbethoughtofasdescribingapointwithina5330dimensionalimagespace.Thissimpleprinciplecaneasilybeextendedtomuchlargerimages:

a256by256pixelimageoccupiesasinglepointin65,536-dimensionalimagespaceandagain,similarimagesoccupyclosepointswithinthatspace.Likewise,similarfacesarelocatedclosetogetherwithintheimagespace,whiledissimilarfacesarespacedfarapart.CalculatingtheEuclideandistanced,betweentwofacialimagevectors（oftenreferredtoasthequeryimageq,andgalleryimageg）,wegetanindicationofsimilarity.Athresholdisthenappliedtomakethefinalverificationdecision.

VerificationTests

Theprimaryconcerninanyfacerecognitionsystemisitsabilitytocorrectlyverifyaclaimedidentityordetermineaperson'smostlikelyidentityfromasetofpotentialmatchesinadatabase.Inordertoassessagivensystem’sabilitytoperformthesetasks,avarietyofevaluationmethodologieshavearisen.Someoftheseanalysismethodssimulateaspecificmodeofoperation（i.e.securesiteaccessorsurveillance）,whileothersprovideamoremathematicaldescriptionofdatadistributioninsomeclassificationspace.Inaddition,theresultsgeneratedfromeachanalysismethodmaybepresentedinavarietyofformats.Throughouttheexperimentationsinthisthesis,weprimarilyusetheverificationtestasourmethodofanalysisandcomparison,althoughwealsouseFisher’sLinearDiscriminatetoanalyzeindividualsubspacecomponentsinsection7andtheidentificationtestforthefinalevaluationsdescribedinsection8.Theverificationtestmeasuresasystem’sabilitytocorrectlyacceptorrejecttheproposedidentityofanindividual.Atafunctionallevel,thisreducestotwoimagesbeingpresentedforcomparison,forwhichthesystemmustreturneitheranacceptance（thetwoimagesareofthesameperson）orrejection（thetwoimagesareofdifferentpeople）.Thetestisdesignedtosimulatetheapplicationareaofsecuresiteaccess.Inthisscenario,asubjectwillpresentsomeformofidentificationatapointofentry,perhapsasaswipecard,proximitychiporPINnumber.Thisnumberisthenusedtoretrieveastoredimagefromadatabaseofknownsubjects（oftenreferredtoasthetargetorgalleryimage）andcomparedwithaliveimagecapturedatthepointofentry（thequeryimage）.Accessisthengranteddependingontheacceptance/rejectiondecision.

Theresultsofthetestarecalculatedaccordingtohowmanytimestheaccept/rejectdecisionismadecorrectly.Inordertoexecutethistestwemustfirstdefineourtestsetoffaceimages.Althoughthenumberofimagesinthetestsetdoesnotaffecttheresultsproduced（astheerrorratesarespecifiedaspercentagesofimagecomparisons）,itisimportanttoensurethatthetestsetissufficientlylargesuchthatstatisticalanomaliesbecomeinsignificant（forexample,acoupleofbadlyalignedimagesmatchingwell）.Also,thetypeofimages（highvariationinlighting,partialocclusionsetc.）willsignificantlyaltertheresultsofthetest.Therefore,inordertocomparemultiplefacerecognitionsystems,theymustbeappliedtothesametestset.

However,itshouldalsobenotedthatiftheresultsaretoberepresentativeofsystemperformanceinarealworldsituation,thenthetestdatashouldbecapturedunderpreciselythesamecircumstancesasintheapplicationenvironment.Ontheotherhand,ifthepurposeoftheexperimentationistoevaluateandimproveamethodoffacerecognition,whichmaybeappliedtoarangeofapplicationenvironments,thenthetestdatashouldpresenttherangeofdifficultiesthataretobeovercome.Thismaymeanincludingagreaterpercentageof‘difficult’imagesthanwouldbeexpectedintheperceivedoperatingconditionsandhencehighererrorratesintheresultsprodu