資源描述：
《最新A Robust Real Time Face Detection：一個(gè)強(qiáng)大的實(shí)時(shí)人臉檢測(cè)教學(xué)講義PPT.ppt》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在教育資源-天天文庫。
1、ARobustRealTimeFaceDetection：一個(gè)強(qiáng)大的實(shí)時(shí)人臉檢測(cè)OutlineAdaBoost–LearningAlgorithmFaceDetectioninreallifeUsingAdaBoostforFaceDetectionImprovementsDemonstrationAdaBoostAshortIntroductiontoBoosting(Freund&Schapire,1999)LogisticRegression,AdaBoostandBregmanDista
2、nces(Collins,Schapire,Singer,2002)AdaBoost–trainingerrorFreundandSchapire(1997)provedthat:AdaBoostADAptstotheerrorratesoftheindividualweakhypotheses.AdaBoost–generalizationerrorFreundandSchapire(1997)showedthat:AdaBoost–generalizationerrorTheanalysis
3、impliesthatboostingwilloverfitifrunfortoomanyroundsHowever,itwasobservedempiricallythatAdaBoostdoesnotoverfit,evenwhenrunthousandsofrounds.Moreover,itwasobservedthatthegeneralizationerrorcontinuestodrivedownlongaftertrainingerrorreachedzeroAdaBoost–g
4、eneralizationerrorAnalternativeanalysiswaspresentedbySchapireetal.(1998),thatsuitstheempiricalfindingsAdaBoost–differentpointofviewWetrytosolvetheproblemofapproximatingthey’susingalinearcombinationofweakhypothesesInotherwords,weareinterestedintheprob
5、lemoffindingavectorofparametersαsuchthat isa‘goodapproximation’ofyiForclassificationproblemswetrytomatchthesignoff(xi)toyiAdaBoost–differentpointofviewSometimesitisadvantageoustominimizesomeother(non-negative)lossfunctioninsteadofthenumberofclassi
6、ficationerrorsForAdaBoostthelossfunctionisThispointofviewwasusedbyCollins,SchapireandSinger(2002)todemonstratethatAdaBoostconvergestooptimalityFaceDetection(notfacerecognition)FaceDetectioninMonkeysTherearecellsthat‘detectfaces’FaceDetectioninHumanTh
7、ereare‘processesoffacedetection’FacesAreSpecialWeanalyzefacesina‘differentway’FacesAreSpecialWeanalyzefacesina‘differentway’FacesAreSpecialWeanalyzefacesina‘differentway’FaceRecognitioninHumanWeanalyzefaces‘inaspecificlocation’RobustReal-TimeFaceDete
8、ctionViolaandJones,2003FeaturesPictureanalysis,IntegralImageFeaturesThesystemclassifiesimagesbasedonthevalueofsimplefeaturesTwo-rectangleThree-rectangleFour-rectangleValue=∑(pixelsinwhitearea)-∑(pixelsinblackarea)ContrastFeaturesSourceResultFeaturesN
