facial feature detection using adaboost with shape constraints

《facial feature detection using adaboost with shape constraints》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、FacialfeaturedetectionusingAdaBoostwithshapeconstraintsDavidCristinacceandTimCootesDept.ImagingScienceandBiomedicalEngineeringUniversityofManchester,Manchester,M139PT,U.K.david.cristinacce@stud.man.ac.ukAbstractRecentlyafastandefficientfacedetectionmethodhasbeen

2、devised[11],whichreliesontheAdaBoostalgorithmandasetofHaarWaveletlikefea-tures.Anaturalextensionofthisapproachistousethesametechniquetolo-cateindividualfeatureswithinthefaceregion.However,wefindthatthereisinsufficientlocalstructuretoreliablylocateeachfeatureineve

3、ryimage,andthuslocalmodelscangivemanyfalsepositiveresponses.Wedemonstratethattheperformanceofsuchfeaturedetectorscanbesignificantlyimprovedbyusingglobalshapeconstraints.Wedescribeanalgorithmcapableofac-curatelyandreliablydetectingfacialfeaturesandpresentquantit

4、ativeresultsonbothhighandlowresolutionimagesets.1IntroductionThispaperaddressestheproblemoflocatingfacialfeatures(eyes,nose,mouthcornersandsoon)inimagesoffrontalfaces.Locatingsuchfeaturesisanimportantstageinmanyfacialimageinterpretationtasks(suchasfaceverificat

5、ion,facetrackingorfaceexpressionrecognition).WeadoptthefastandefficientfacefinderrecentlydescribedbyViolaandJones[11]tolocatetheapproximatepositionofeachfaceinanimage.Wethenusethesamemethod,trainedonregionsaroundfacialfeaturepoints,tolocateinteriorpointsontheface

6、.However,thereisofteninsufficientlocalstructurearoundeachfeaturetotrainreallyreliablefeaturefinders.Wefindthatwhensetwiththresholdssufficienttolocatethetruepositionreasonablyfrequently,suchdetectorsproducemanyfalsepositives.Toselectthemostsuitablecandidatesweusestat

7、isticalmodelsoftheconfigurationsofthepoints.Wefindthatcombiningfeaturedetectorswithsuchstatisticalshapemodelsgivesasignificantimprovementinboththereliabilityandtheoverallaccuracyofthefeaturedetectionsystem.Inthefollowingwedescribetheapproachinmoredetail,anddemons

8、trateitsappli-cationtofindingfeaturesintwodatasets.Thoughdemonstratedonfaces,theapproachisclearlyapplicabletoawidevarietyofimageinterpretationtasks.2BackgroundFacedetectionhasreceiv