资源描述:
《Deep Learning for Understanding Faces英文文献资料》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、DEEPLEARNINGFORVISUALUNDERSTANDING:PART2RajeevRanjan,SwamiSankaranarayanan,AnkanBansal,NavaneethBodla,Jun-ChengChen,VishalM.Patel,CarlosD.Castillo,andRamaChellappaDeepLearningforUnderstandingFacesMachinesmaybejustasgood,orbetter,thanhumansecentdevelopmentsindeepconvolutionalneuralnet
2、works(DCNNs)haveshownimpressiveperformanceimprove-Rmentsonvariousobjectdetection/recognitionproblems.Thishasbeenmadepossibleduetotheavailabilityoflargeanno-tateddataandabetterunderstandingofthenonlinearmappingbetweenimagesandclasslabels,aswellastheaffordabilityofpowerfulgraphicsproce
3、ssingunits(GPUs).Thesedevelopmentsindeeplearninghavealsoimprovedthecapabilitiesofmachinesinunderstandingfacesandautomaticallyexecutingthetasksoffacedetection,poseestimation,landmarklocalization,andfacerecognitionfromunconstrainedimagesandvideos.Inthisarticle,weprovideanoverviewofdeep
4、-learningmethodsusedforfacerecognition.Wediscussdifferentmodulesinvolvedindesigninganautomaticfacerecognitionsystemandtheroleofdeeplearningforeachofthem.SomeopenissuesregardingDCNNsforfacerecognitionproblemsarethendiscussed.Thisarticleshouldprovevaluabletoscientists,engineers,andendu
5、sersworkinginthefieldsoffacerecognition,security,visualsurveillance,andbiometrics.WhatwecanlearnfromfacesFacialanalyticsisachallengingproblemincomputervisionandhasbeenactivelyresearchedformorethantwodecades[1].Thegoalistoextractasmuchinformationaspossible,suchasloca-tion,pose,gender,
6、ID,age,emotion,andsoforthfromaface.©ISTOCKPHOTO.COM/ZAPP2PHOTOApplicationsofthistechnologyincludedetectingandidentifyingapersonofinterestfromsurveillancevideos,activeauthentica-tionofusers’cellphones,paymenttransactionsusingfacebio-metrics,smartcars,etc.Inaddition,therehasbeenagrowin
7、ginterestinfacerecognitionandverificationfromunconstrainedimagesandvideosthatalsoinvolvesubtaskssuchasfacedetec-tion,faciallandmarklocalization,etc.Inthisarticle,wepresentanoverviewofrecentautomaticfaceidentificationandverificationsystemsbasedondeeplearning.Threemodulesaretypicallyne
8、ededforsucha