资源描述:
《Recursive Spatial Transformer (ReST) for Alignment-Free Face Recognition》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、RecursiveSpatialTransformer(ReST)forAlignment-FreeFaceRecognitionWanglongWu1,2MeinaKan1,3XinLiu1,2YiYang4ShiguangShan1,3XilinChen11KeyLabofIntelligentInformationProcessingofChineseAcademyofSciences(CAS),InstituteofComputingTechnology,CAS,Beijing100190,China2U
2、niversityofChineseAcademyofSciences,Beijing100049,China3CASCenterforExcellenceinBrainScienceandIntelligenceTechnology4HuaweiTechnologiesCo.,Ltd.,Beijing100085,China{wanglong.wu,meina.kan,xin.liu,shiguang.shan,xilin.chen}@vipl.ict.ac.cnyangyi16@huawei.comAbstr
3、actConvolutionalNeuralNetwork(CNN)hasledtosignif-icantprogressinfacerecognition.CurrentlymostCNN-basedfacerecognitionmethodsfollowatwo-steppipeline,Figure1.Conventionalfacerecognitionpipelinei.e.adetectedfaceisfirstalignedtoacanonicalonepre-definedbyameanfacesh
4、ape,andthenitisfedintoapowerfuldeeplearningmodels,particularlyConvolutionalCNNtoextractfeaturesforrecognition.ThealignmentstepNeuralNetwork(CNN),andhasmetprominentboostoftransformsallfacestothesameshape,whichcancauserecognitionaccuracy[16,14,17].Thoughequippe
5、dwithlossofgeometricalinformationwhichishelpfulindistin-newmodels,mostCNN-basedfacerecognitionapproachesguishingdifferentsubjects.Moreover,itishardtodefinestillfollowtheconventionalrecognitionpipeline.Asillus-asingleoptimalshapeforthefollowingrecognition,since
6、tratedinFig.1,givenaninputimage,facedetectionisfirstfaceshavelargediversityinfacialfeatures,e.g.poses,illu-performedtoobtaintheboundingboxofeachface.Thenmination,etc.Tobefreefromtheaboveproblemswithanthedetectedfacesarealignedtoacanonicaloneformoreindependenta
7、lignmentstep,weintroduceaRecursiveSpa-robustfeatureextraction,andfinallythealignedfacesaretialTransformer(ReST)moduleintoCNN,allowingfaceusedtorecognizethesubjects.alignmenttobejointlylearnedwithfacerecognitioninanInthestepoffacealignment,detectedfacesaretrans
8、-end-to-endfashion.ThedesignedReSThasanintrinsicre-formedtoacanonicaloneaccordingtotheaffinerela-cursivestructureandiscapableofprogressivelyaligningtionshipsbetweentheirfaciallandmarksanda