资源描述:
《SSPP-DAN -- Deep Domain Adaptation Network for Face Recognition with Single Sample Per Person 英文文献资料》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、SSPP-DAN:DEEPDOMAINADAPTATIONNETWORKFORFACERECOGNITIONWITHSINGLESAMPLEPERPERSONSungeunHong,WoobinIm,JongbinRyu,HyunS.YangSchoolofComputing,KoreaAdvancedInstituteofScienceandTechnology,RepublicofKoreaABSTRACTReal-worldfacerecognitionusingasinglesampleperper
2、son(SSPP)isachallengingtask.Theproblemisexacerbatediftheconditionsunderwhichthegalleryimageandtheprobesetarecapturedarecompletelydifferent.Toaddresstheseis-suesfromtheperspectiveofdomainadaptation,weintroduceanSSPPdomainadaptationnetwork(SSPP-DAN).Inthepro
3、-posedapproach,domainadaptation,featureextraction,and(a)(b)(c)classificationareperformedjointlyusingadeeparchitectureFig.1:Examplesof(a)astablegalleryimage(sourcedo-withdomain-adversarialtraining.However,theSSPPcharac-main)(b)syntheticimagesgeneratedtooverc
4、omethelackofteristicofonetrainingsampleperclassisinsufficienttotraingallerysamples(sourcedomain)(c)unstableprobeimagesthedeeparchitecture.Toovercomethisshortage,wegeneratethatincludeblur,noise,andposevariation(targetdomain)syntheticimageswithvaryingposesusi
5、nga3Dfacemodel.ExperimentalevaluationsusingarealisticSSPPdatasetshowinDA,amappingbetweenthesourcedomainandthetar-thatdeepdomainadaptationandimagesynthesiscomplementgetdomainisconstructed,suchthattheclassifierlearnedforeachotheranddramaticallyimproveaccuracy
6、.Experimentsthesourcedomaincanalsobeappliedtothetargetdomain.onabenchmarkdatasetusingtheproposedapproachshowInspiredbythis,weassumestableshootingconditionofastate-of-the-artperformance.gallerysetasthesourcedomainandunstableshootingcon-ditionofaprobesetasth
7、etargetdomainasshowninFig.1.IndexTerms—SSPPfacerecognition,One-shotlearn-ToapplyDAintheunifieddeeparchitecture,weuseadeeping,Unsuperviseddomainadaptation,Faceimagesynthesis,neuralnetworkwithdomain-adversarialtraining,inamannerSurveillancecameraproposedin[3]
8、.Thebenefitofthisapproachisthatlabelsinthetargetdomainarenotrequiredfortraining,i.e.,theap-1.INTRODUCTIONproachaccommodatesunsupervisedlearning.ThesecondchallengeinusingSSPPisintheshortageofThereareseveralexam