9 A New Feature Set for Face Detection台湾清华大学的一篇硕士毕业论文全英文较详细吧.docx
《9 A New Feature Set for Face Detection台湾清华大学的一篇硕士毕业论文全英文较详细吧.docx》由会员分享，可在线阅读，更多相关《9 A New Feature Set for Face Detection台湾清华大学的一篇硕士毕业论文全英文较详细吧.docx（38页珍藏版）》请在冰点文库上搜索。
9ANewFeatureSetforFaceDetection台湾清华大学的一篇硕士毕业论文全英文较详细吧
國立清華大學
碩士論文
題目:
使用新特徵的人臉偵測系統
ANewFeatureSetforFaceDetection
系別資訊系統與應用研究所組別甲
學號姓名926701蔡忠志（Chung-ChihTsai）
指導教授張智星博士（Jyh-ShingRogerJang）
中華民國九十四年六月
Abstract
ViolaandJonesintroduceafastfacedetectionsystemwhichusesacascadedstructurethatcanachievehighdetectionrateandlowfalsepositiverate.Theirsystemusesintegralimagestocomputevaluesfromfeatures.Thisthesisintroducestwonewtypesofintegralimageswhicharecalledtriangleintegralimagesandtwocorrespondingfeatureswhicharenamedtrianglefeatures.AndthisthesisproposesamethodtolowertrainingerrorbymodifyingDiscreteAdaBoost.Asresults,tousetrianglefeaturescandecreasethenumbersoffeatures;thisresearchachieveslowerfalsepositiverateandfewerfeaturesareused.
摘要
在快速人臉偵測的研究中，Viola和Jones提出了一個連接式的架構，此架構能得到高辨識率及低錯誤率；他們使用integralimage來計算人臉特徵值。
本研究提出了兩種新的integralimage:
triangleintegralimage以及各別對應的三角特徵。
另外，本研究以DiscreteAdaBoost為基礎，提出了一個能在訓練時降低非人臉錯誤率的方法。
我們的實驗證明，三角特徵能使得需要的feature數減少；改進過後的AdaBoost能使得錯誤率更低。
Keywords
Facedetection,integralimage,feature,AdaBoost,cascadestructure
TABLEOFCONTENTS
Abstracti
Keywordsi
Acknowledgementiv
1Introduction1
1.1SystemOverview1
1.2ThesisOrganization2
2RelatedWork3
3IntegralImagesandFeatures5
3.1IntegralImages5
3.1.1RectangleIntegralImage（RII）5
3.1.2TriangleIntegralImages6
3.2Features8
3.2.1ExtensionofRectangleFeatures9
3.2.2TriangleFeatures10
4LearningStrongClassifiers13
4.1WeakClassifier13
4.2LearnaStrongClassifier14
4.2.1DiscreteAdaBoost（DA）14
4.2.2DiscussionaboutDA17
4.2.3ModificationstoDA18
4.2.4DiscussionaboutModifications18
5CascadeofStrongClassifiers20
5.1LearningData20
5.2LearnaCascadedClassifier21
6ExperimentalResults23
6.1Imageprocessing23
6.2ScanningImages27
6.3ExperimentalResults27
6.3.1Dataset28
6.3.2SelectionofFeatures29
6.3.3SystemPerformance32
6.3.4ErrorAnalyses34
7ConclusionsandFutureWork36
Referencesv
AppendixA:
SamplesofDetectionResultsvi
AppendixB:
ListofFeatureTypesSelectedinEachStagevii
LISTOFFIGURES
Figure1.1:
Flowchartofthefacedetectionsystem1
Figure3.1:
Arectangleintegralimage5
Figure3.2:
TriangleIntegralImage17
Figure3.3:
TriangleIntegralImage27
Figure3.4:
Facecharacteristicsproposedin[1][2]8
Figure3.5:
Featuretypesproposedin[3]8
Figure3.6:
Extendedrectanglefeaturetypes9
Figure3.7:
Calculationofrectanglefeature10
Figure3.8:
Computethesumofpixelsinarectanglearea10
Figure3.9:
Type7:
TriangleFeatureType111
Figure3.10:
Type8:
TriangleFeatureType211
Figure3.11:
Type9:
combinationoffeatures12
Figure4.1:
Thresholdsofaweakclassifier13
Figure4.2:
Samplesofre-weightingprocess17
Figure5.1:
Cascadedstructure20
Figure6.1:
Animageresizingsample24
Figure6.2:
ContrastStretching25
Figure6.3:
Examplesofimageprocessing26
Figure6.4:
Samplesoffaceandnon-faceimages29
Figure6.5:
Samplesoftype7andtype829
Figure6.6:
Totalnumbersoffeaturesofeachtype30
Figure6.7:
Comparisonoftrainingerrorofusingnewfeaturetypes31
Figure6.8:
FPRofeachstage32
Figure6.9:
ComparisonoftrainingerrorofmodificationtoGA33
Figure6.10:
Anexampleofmisclassification34
LISTOFTABLES
Table4.1:
ProcedureofDiscreteAdaBoost15
Table5.1:
Cascadedclassifierlearningalgorithm22
Table6.1:
Numbersoffeaturesofeachtype30
Table6.2:
Comparisonofperformance34
Acknowledgement
在清華的兩年中，首先要衷心感謝指導教授張智星老師的指導，無論在做人處事或專業領域上的啟發都讓我獲益良多，使我能順利完成本篇論文；並且感謝口試委員的指導，使本論文更加完善。
另外要感謝MIR實驗室的各位，有和大家的互相砥礪、創意的激發，才使得研究生活不致枯燥乏味。
感謝我的家人這兩年來的支持及關心。
最後要感謝大學的同窗好友jclin，這幾年的生活真的很有意思。
1Introduction
Facedetectionisanimportantcomponentofacontent-basedvideoinformationretrievalsystem.
Therearethreemaindirectionsforfastfacedetectionresearches.Thefirstoneistofindnewusefulfeaturetypestodecreasethenumberofclassifiers.Thesecondoneistomodifyexistentlearningprocessorintroduceanewonetoselectfeatures.Thethirdoneistodecreasethenumbersofsub-imagestodetecttospeedupthedetectionspeed.Thisresearchfocusesonthefirsttwodirections.
Thisresearchhastwomaincontributions.First,weintroducetwonovelkindsofintegralimagesandfeaturetypes.Second,weobservesomeproblemsofDiscreteAdaBoostandmodifythelearningalgorithm.Thisresearchfocusesondetectionsofupright-frontalfaces.
1.1SystemOverview
Theflowchartofoursystemisshownasfigure1.1.
Figure1.1:
Flowchartofthefacedetectionsystem
Thefacedetectioncomponentisacascadedstructure.Thestructureiscascadedbystrongclassifiers.Eachstrongclassifierconsistsofseveralweakclassifiers.Andaweakclassifierconsistsofaweightandafeaturewiththresholds.
Wecangeneratemanysub-imagesfromanimagebyvariouspositionsandscales.Eachstrongclassifierrejectsnumbersofsub-images;therejectedsub-imagesarenolongerbeingprocessed.Mostofthoserejectedsub-windowsarenon-faceimages,andfewofthemarefaceimages.
Wehavetodefinedetectionrate（DR）andfalsepositive（FP）first.DRistheratioofnumberoffaceimageswhicharecorrectlydetectedtototalfacenumber,e.g.80facesaredetectedoutof100faces,DR=0.8.FPisanumberofnon-faceimageswhicharedetectedasfaceimages;therateofFPtototalnon-facesisthereforecalledfalsepositiverate（FPR）.
1.2ThesisOrganization
Thisthesisisorganizedasfollows:
chaptertwointroducestherelatedfacedetectionresearches;chapterthreeintroducesintegralimagesandfeaturesusedinoursystem;chapterfourintroducesalearningalgorithmtoselectfeaturesandtrainastrongclassifier;chapterfiveintroducesalearningprocesstotrainacascadedclassifier;chaptersixshowstheexperimentsandresults;chapterseventalksabouttheconclusionsandfutureworks.
2RelatedWork
In2001,ViolaandJonesintroducearapidobjectdetectionsystem[1][2].Theirresearchhasthreemaincontributions.First,theyintroduceintegralimagewhichallowsfastcomputationoffeatures.Second,theyuseAdaBoost[4]totrainefficientclassifiers.Third,theyintroduceacascadedstructurewhichcanrejectnon-faceimagesquickly.Theirsystemcanachievehighdetectionratewithsmallnumberoffalsepositives.
Laterin2002,Lienhartetal.extendViola’sresearchandtheirresearchhasthreemaincontributions[3].First,theyintroduceanovelfeaturesetwhichisdesignedfordetectingin-planerotationfaces.Second,theypresentanalysesamongthedifferentboostingalgorithms（Discrete,RealandGentleAdaBoost）.Third,theycomparetheperformancebetweenstumpsandRegressionTree（CART）andalsoanalyzetheeffectofsizesoftrainingdata.
StanZ.LiandZhenQiuZhangintroduceanovellearningprocedurewhichiscalledFloatBoost[5].FloatBoostcomesfromthefloatingsearchalgorithm.Recallthattherearebasicallythreekindsoffeatureselectionmethods:
SequentialForwardSelection（SFS）whichisusedinAdaBoost,SequentialBackwardSelection（SBS）andSequentialFloatingSearchMethod（SFSM）whichcombinesSFSandSBS.SFSMcanachieveapproximateoptimalcombinationofselection.FloatBoostusesSFSMtoselectfeatures;thetrainingtimeisfivetimeslongerthanAdaBoost.Theyalsointroduceapyramidstructurefordetectingmultipleout-of-planedegreefaces.
YongMaandXiaoqingDingintroduceCostSensitive-AdaBoost（CS-AdaBoost）[6].TheweaklearnercanselectmorefeaturesbyusingCS-AdaBoost.WeproposeasimilarmodificationtoAdaBoostforselectingmorefeaturesalso.
DongZhangetal.introduceafacedetectionframeworkwhichusesdifferentkindsoffeaturesinearlystagesandlatestagesbecauselocalfeatures（haar-likefeatures）maynotbeveryusefulinlatestages[7].Thustheyuseglobalfeatureinlatestages.GlobalfeatureusesPCA（PrincipalComponentAnalysis）features.
Intheintroduction,webringupthedirectionsoffacedetectionresearches.Assummary,[1][2]introduceaframeworkoffastfacedetectionsystem,[3][7]introducenewfeaturetypesand[5][6]modifythetrainingprocess.
3IntegralImagesandFeatures
Featuresusedbyoursystemcanbecomputedveryfastthroughintegralimages.Inthischapter,wewillintroduceintegralimages,featuretypesandthewaytouseintegralimagestocalculateavaluefromafeature.
3.1IntegralImages
IntegralimageisalsocalledSummedAreaTable（SAT）.Itrepresentsasumofaparticularareainanimage.
3.1.1RectangleIntegralImage（RII）
RIIisintroducedin[1][2].Thevalueatposition（x,y）inaRIIrepresentsthesumofpixelsaboveandleftto（x,y）intheoriginalimage:
whereRII（x,y）isthevalueofRIIatposition（x,y）andI（x’,y’）isthepixelvalueoftheoriginalimageatposition（x’,y’）.
（x,y）
Figure3.1:
Arectangleintegralimage
Foreachimage,wecomputeitsRIIthroughonlyonepassofscanningthepixelsinanimage.Inpractice,firstwecomputethecumulativerowsumandthenaddthesumtoRIIatpreviousrowandthesamecolumntogetthesumofpixelsaboveandleft:
（3-1）
（3-2）
DuringcomputingaRII,wealsocomputethesquaresumoftheimageforcalculatingitsvarianceforcontraststretching;section6.1hasdetaileddescriptions.
3.1.2TriangleIntegralImages
ThisresearchintroducestwonewtypesofintegralimageswhicharecalledTriangleIntegralImages（TIIs）.TheideaofTIIscomesfromtheRotatedSummedAreaTable（RSAT）in[3].TIIsrepresentthesumsofrighttriangleareasinanimage.T
升级会员 