人脸识别文献翻译中英双文复习课程文档格式.docx
- 文档编号:14849839
- 上传时间:2022-10-25
- 格式:DOCX
- 页数:9
- 大小:101.39KB
人脸识别文献翻译中英双文复习课程文档格式.docx
《人脸识别文献翻译中英双文复习课程文档格式.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复习课程文档格式.docx(9页珍藏版)》请在冰豆网上搜索。
人脸识别文献翻译中英双文复习课程@#@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’toencompassalltechn
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- 识别 文献 翻译 中英双文 复习 课程
![提示](https://static.bdocx.com/images/bang_tan.gif)