人脸识别文献翻译中英文讲课讲稿.docx

人脸识别文献翻译中英文讲课讲稿.docx

《人脸识别文献翻译中英文讲课讲稿.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文讲课讲稿.docx（19页珍藏版）》请在冰豆网上搜索。

人脸识别文献翻译中英文讲课讲稿

附录（原文及译文）

翻译原文来自

ThomasDavidHeseltineBSc.Hons.TheUniversityofYork

DepartmentofComputerScience

FortheQualificationofPhD.--September2005-

《FaceRecognition:

Two-DimensionalandThree-DimensionalTechniques》

4Two-dimensionalFaceRecognition

4.1FeatureLocalization

Beforediscussingthemethodsofcomparingtwofacialimageswenowtakeabrieflookatsomeatthepreliminaryprocessesoffacialfeaturealignment.Thisprocesstypicallyconsistsoftwostages:

facedetectionandeyelocalisation.Dependingontheapplication,ifthepositionofthefacewithintheimageisknownbeforehand（foracooperativesubjectinadooraccesssystemforexample）thenthefacedetectionstagecanoftenbeskipped,astheregionofinterestisalreadyknown.Therefore,wediscusseyelocalisationhere,withabriefdiscussionoffacedetectionintheliteraturereview（section3.1.1）.

Theeyelocalisationmethodisusedtoalignthe2Dfaceimagesofthevarioustestsetsusedthroughoutthissection.However,toensurethatallresultspresentedare

representativeofthefacerecognitionaccuracyandnotaproductoftheperformanceoftheeyelocalisationroutine,allimagealignmentsaremanuallycheckedandanyerrorscorrected,priortotestingandevaluation.

Wedetectthepositionoftheeyeswithinanimageusingasimpletemplatebased

method.Atrainingsetofmanuallypre-alignedimagesoffacesistaken,andeach

imagecroppedtoanareaaroundbotheyes.Theaverageimageiscalculatedandused

asatemplate.

Figure4-1-Theaverageeyes.Usedasatemplateforeyedetection.

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

Awindowispassedoverthetestimagesandtheabsolutedifferencetakentothatoftheaverageeyeimageshownabove.Theareaoftheimagewiththelowestdifferenceistakenastheregionofinterestcontainingtheeyes.Applyingthesameprocedureusingasmallertemplateoftheindividualleftandrighteyesthenrefineseacheyeposition.

Thisbasictemplate-basedmethodofeyelocalisation,althoughprovidingfairlypreciselocalisations,oftenfailstolocatetheeyescompletely.However,weareableto

improveperformancebyincludingaweightingscheme.

Eyelocalisationisperformedonthesetoftrainingimages,whichisthenseparatedintotwosets:

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

Figure4-2–Distancetotheeyetemplateforsuccessfuldetections（top）indicatingvariancedueto

noiseandfaileddetections（bottom）showingcrediblevarianceduetomiss-detectedfeatures.

Inthelowerimage（Figure4-2bottom）,wehavetakenthesetoffailedlocalisations（imagesoftheforehead,nose,cheeks,backgroundetc.falselydetectedbythelocalisationroutine）andonceagaincomputedtheaveragedistancefromtheeyetemplate.Thebrightpupilssurroundedbydarkerareasindicatethatafailedmatchisoftenduetothehighcorrelationofthenoseandcheekboneregionsoverwhelmingthepoorlycorrelatedpupils.Wantingtoemphasisethedifferenceofthepupilregionsforthesefailedmatchesandminimisethevarianceofthewhitesoftheeyesforsuccessfulmatches,wedividethelowerimagevaluesbytheupperimagetoproduceaweightsvectorasshowninFigure4-3.Whenappliedtothedifferenceimagebeforesummingatotalerror,thisweightingschemeprovidesamuchimproveddetectionrate.

Figure4-3-Eyetemplateweightsusedtogivehigherprioritytothosepixelsthatbestrepresenttheeyes.

4.2TheDirectCorrelationApproach

Webeginourinvestigationintofacerecognitionwithperhapsthesimplestapproach,knownasthedirectcorrelationmethod（alsoreferredtoastemplatematchingbyBrunelliandPoggio[29]）involvingthedirectcomparisonofpixelintensityvaluestakenfromfacialimages.Weusetheterm‘DirectCorrelation’toencompassalltechniquesinwhichfaceimagesarecompareddirectly,withoutanyformofimagespaceanalysis,weightingschemesorfeatureextraction,regardlessofthedistancemetricused.Therefore,wedonotinferthatPearson’scorrelationisappliedasthesimilarityfunction（althoughsuchanapproachwouldobviouslycomeunderourdefinitionofdirectcorrelation）.WetypicallyusetheEuclideandistanceasourmetricintheseinvestigations（inverselyrelatedtoPearson’scorrelationandcanbeconsideredasascaleandtranslationsensitiveformofimagecorrelation）,asthispersistswiththecontrastmadebetweenimagespaceandsubspaceapproachesinlatersections.

Firstly,allfacialimagesmustbealignedsuchthattheeyecentresarelocatedattwospecifiedpixelcoordinatesandtheimagecroppedtoremoveanybackground

information.Theseimagesarestoredasgreyscalebitmapsof65by82pixelsandpriortorecognitionconvertedintoavectorof5330elements（eachelementcontainingthecorrespondingpixelintensityvalue）.Eachcorrespondingvectorcanbethoughtofasdescribingapointwithina5330dimensionalimagespace.Thissimpleprinciplecaneasilybeextendedtomuchlargerimages:

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

dqg（dthreshold⇒accept）（dthreshold⇒reject）.Equ.4-1

4.2.1VerificationTests

Theprimaryconcerninanyfacerecognitionsystemisitsabilitytocorrectlyverifyaclaimedidentityordetermineaperson'smostlikelyidentityfromasetofpotentialmatchesinadatabase.Inordertoassessagivensystem’sabilitytoperformthesetasks,avarietyofevaluationmethodologieshavearisen.Someoftheseanalysismethodssimulateaspecificmodeofoperation（i.e.securesiteaccessorsurveillance）,whileothersprovideamoremathematicaldescriptionofdatadistributioninsome

classificationspace.Inaddition,theresultsgeneratedfromeachanalysismethodmay

bepresentedinavarietyofformats.Throughouttheexperimentationsinthisthesis,weprimarilyusetheverificationtestasourmethodofanalysisandcomparison,althoughwealsouseFisher’sLinearDiscriminanttoanalyseindividualsubspacecomponentsinsection7andtheidentificationtestforthefinalevaluationsdescribedinsection8.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,ifthepurposeoftheexperimentationistoevaluateandimproveametho