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（16页珍藏版）》请在冰点文库上搜索。

人脸识别文献翻译中英双文

4Two-dimensionalFaceRecognition

4.1FeatureLocalization

Beforediscussingthemethodsofcomparingtwofacial

theimageisknownbeforehand（foracooperativesubjectinadooraccesssystemforexample）thenthefacedetectionstagecanoftenbeskipped,as

theregionofinterestisalreadyknown.Therefore,wediscusseye

localizationhere,withabriefdiscussionoffacedetectionintheliteraturereview.

Theeyelocalizationmethodisusedtoalignthe2Dfaceimagesofthevarioustestsetsusedthroughoutthissection.However,toensurethatallresultspresentedarerepresentativeofthefacerecognitionaccuracyandnot

aproductoftheperformanceoftheeyelocalizationroutine,allimagealignmentsaremanuallycheckedandanyerrorscorrected,priortotestingandevaluation.

Wedetectthepositionoftheeyeswithinanimageusingasimpletemplatebasedmethod.Atrainingsetofmanuallypre-alignedimagesoffacesistaken,

andeachimagecroppedtoanareaaroundbotheyes.Theaverageimageiscalculatedandusedasatemplate.

Botheyesareincludedinasingletemplate,ratherthanindividuallysearchingforeacheyeinturn,asthecharacteristicsymmetryoftheeyeseithersideofthenose,provideausefulfeaturethathelpsdistinguishbetweentheeyesandotherfalsepositivesthatmaybepickedupinthebackground.Althoughthismethodishighlysusceptibletoscale（i.e.subjectdistancefromthecamera）andalsointroducestheassumptionthateyesintheimageappearnearhorizontal.Somepreliminaryexperimentationalsorevealsthatitisadvantageoustoincludetheareaofskinjustbeneaththeeyes.

精品文档

Thereasonbeingthatinsomecasestheeyebrowscancloselymatchthetemplate,particularlyifthereareshadowsintheeye-sockets,buttheareaofskinbelowtheeyeshelpstodistinguishtheeyesfromeyebrows（theareajustbelowtheeyebrowscontaineyes,whereastheareabelowtheeyescontainsonlyplainskin）.

Awindowispassedoverthetestimagesandtheabsolutediffereneetakentothatoftheaverageeyeimageshownabove.Theareaoftheimagewiththelowestdifferenceistakenastheregionofinterestcontainingtheeyes.

Applyingthesameprocedureusingasmallertemplateoftheindividualleftandrighteyesthenrefineseacheyeposition.

Thisbasictemplate-basedmethodofeyelocalization,althoughprovidingfairlypreciselocalizations,oftenfailstolocatetheeyescompletely.

However,weareabletoimproveperformancebyincludingaweightingscheme.

Eyelocalizationisperformedonthesetoftrainingimages,whichisthenseparatedintotwosets:

thoseinwhicheyedetectionwassuccessful;andthoseinwhicheyedetectionfailed.Takingthesetofsuccessfullocalizationswecomputetheaveragedistancefromtheeyetemplate（Figure4-2top）.Notethattheimageisquitedark,indicatingthatthedetectedeyescorrelateclosely

totheeyetemplate,aswewouldexpect.However,brightpointsdooccurnearthewhitesoftheeye,suggestingthatthisareaisofteninconsistent,varyinggreatlyfromtheaverageeyetemplate.

Figure4-2-Distancetotheeyetemplateforsuccessfuldetections（top）indicating

varianceduetonoiseandfaileddetections（bottom）showingcrediblevariancedueto

miss-detectedfeatures.

Inthelowerimage（Figure4-2bottom）,wehavetakenthesetoffailedlocalizations（imagesoftheforehead,nose,cheeks,backgroundetc.falselydetectedbythelocalizationroutine）andonceagaincomputedtheaveragedistancefromtheeyetemplate.Thebrightpupilssurroundedbydarkerareasindicatethatafailedmatchisoftenduetothehighcorrelationofthenoseandcheekboneregionsoverwhelmingthepoorlycorrelatedpupils.Wantingtoemphasizethedifferenceofthepupilregionsforthesefailedmatchesand

2欢迎下载

精品文档

minimizethevarianeeofthewhitesoftheeyesforsuccessfulmatches,wedividethelowerimagevaluesbytheupperimagetoproduceaweightsvectorasshowninFigure4-3.Whenappliedtothedifferenceimagebeforesummingatotalerror,thisweightingschemeprovidesamuchimproveddetectionrate.

Figure4-3-Eyetemplateweightsusedtogivehigherprioritytothosepixelsthatbestrepresenttheeyes.

4.2TheDirectCorrelationApproach

Webeginourinvestigationintofacerecognitionwithperhapsthesimplest

approach,knownasthedirectcorrelationmethod（alsoreferredtoastemplatematchingbyBrunelliandPoggio）involvingthedirectcomparisonofpixelintensityvaluestakenfromfacialimages.WeusethetermDirectCorrelation'

toencompassalltechniquesinwhichfaceimagesarecompareddirectly,withoutanyformofimagespaceanalysis,weightingschemesorfeature

extraction,regardlessofthedistancemetricused.Therefore,wedonotinfer

thatPearson'scorrelationisappliedasthesimilarityfunction（althoughsuchanapproachwouldobviouslycomeunderourdefinitionofdirect

correlation）.WetypicallyusetheEuclideandistanceasourmetricintheseinvestigations（inverselyrelatedtoPearson'scorrelationandcanbe

consideredasascaleandtranslationsensitiveformofimagecorrelation）,asthispersistswiththecontrastmadebetweenimagespaceandsubspaceapproachesinlatersections.

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

a256by256pixelimageoccupiesasinglepointin65,536-dimensionalimagespaceandagain,similarimagesoccupyclosepoints

withinthatspace.Likewise,similarfacesarelocatedclosetogetherwithin

3欢迎下载

精品文档

theimagespace,whiledissimilarfacesarespacedfarapart.CalculatingtheEuclideandistanced,betweentwofacialimagevectors（oftenreferredtoasthequeryimageq,andgalleryimageg）,wegetanindicationofsimilarity.

Athresholdisthenappliedtomakethefinalverificationdecision.

4.2.1VerificationTests

Theprimaryconcerninanyfacerecognitionsystemisitsabilitytocorrectlyverifyaclaimedidentityordetermineaperson'smostlikelyidentityfromasetofpotentialmatchesinadatabase.Inordertoassessagivensystem'sabilitytoperformthesetasks,avarietyofevaluationmethodologieshavearisen.Someoftheseanalysismethodssimulateaspecificmodeofoperation（i.e.securesiteaccessorsurveillance）,whileothersprovideamoremathematicaldescriptionofdatadistributioninsomeclassificationspace.Inaddition,theresultsgeneratedfromeachanalysismethodmaybepresentedinavarietyofformats.Throughouttheexperimentationsinthisthesis,weprimarilyusetheverificationtestasourmethodofanalysisandcomparison,althoughwealsouseFisher'sLinear

Discriminatetoanalyzeindividualsubspacecomponentsinsection7andtheidentificationtestforthefinalevaluationsdescribedinsection8.Theverificationtestmeasuresasystem'sabilitytocorrectlyacceptorreject

theproposedidentityofanindividual.Atafunctionallevel,thisreducestotwoimagesbeingpresentedforcomparison,forwhichthesystemmustreturneitheranacceptance（thetwoimagesareofthesameperson）orrejection（thetwoimagesareofdifferentpeople）.Thetestisdesignedtosimulatetheapplicationareaofsecuresiteaccess.Inthisscenario,asubjectwillpresentsomeformofidentificationatapointofentry,perhapsasaswipecard,proximitychiporPINnumber.Thisnumberisthenusedtoretrieveastoredimagefromadatabaseofknownsubjects（oftenreferredtoasthetargetorgalleryimage）andcomparedwithaliveimagecapturedatthepointofentry（thequeryimage）.Accessisthengranteddependingontheacceptance/rejectiondecision.

Theresultsofthetestarecalculatedaccordingtohowmanytimestheaccept/rejectdecisionismadecorrectly.Inordertoexecutethistestwemustfirstdefineourtestsetoffaceimages.Althoughthenumberofimagesinthetestsetdoesnotaffecttheresultsproduced（astheerrorratesare

4欢。

迎下载

精品文档

specifiedaspercentagesofimagecomparisons）,itisimportanttoensurethatthetestsetissufficientlylargesuchthatstatisticalanomaliesbecomeinsignificant（forexample,acoupleofbadlyalignedimagesmatchingwell）.Also,thetypeofimages（highvariationinlighting,partialocclusionsetc.）willsignificantlyaltertheresultsofthetest.Therefore,inordertocomparemultiplefacerecognitionsystems,theymustbeappliedtothesametestset.

However,itshouldalsobenotedthatiftheresultsaretoberepresentativeofsystemperformanceinarealworldsituation,thenthetestdatashouldbecapturedunderpreciselythesamecircumstancesasintheapplicationenvironment.Ontheotherhand,ifthepurposeoftheexperimentationistoevaluateandimproveamethodoffacerecognition,whichmaybeappliedtoarangeofapplicationenvironments,thenthetestdatashouldpresenttherangeofdifficultiesthataretobeovercome.Thismaymeanincludingagreaterpercentageof‘difficult'imagesthanwouldbeexpectedintheperceivedoperatingconditionsandhencehighererrorratesintheresultsproduced.Belowweprovidethealgorithmforexecutingtheverificationtest.Thealgorithmisappliedtoasingletestsetoffaceimages,