搜档网
当前位置:搜档网 › 基于分布式低秩表示的子空间聚类算法

基于分布式低秩表示的子空间聚类算法

计算机研究与发展DOI:10.7544桙issn1000‐1239.2016.20148362JournalofComputerResearchandDevelopment53(7):16051611,2016 收稿日期:2014-12-09;修回日期:2015-06-09

 基金项目:国家自然科学基金项目(61373055);江苏省自然科学基金项目(BK20140419);江苏省高校自然科学研究计划重大项目

(14KJB520001)

ThisworkwassupportedbytheNationalNaturalScienceFoundationofChina(61373055),theNaturalScienceFoundationofJiangsuProvinceofChina(BK20140419),andtheMajorProgramofResearchPlanoftheNaturalScienceinHigherEducationofJiangsuProvinceofChina(14KJB520001). 通信作者:吴小俊(wu_xiaojun@jiangnan.edu.cn)基于分布式低秩表示的子空间聚类算法

许 凯 吴小俊 尹贺峰

(江南大学物联网工程学院 江苏无锡 214122)

(xukai347@sina.com)

DistributedLowRankRepresentation‐BasedSubspaceClusteringAlgorithmXuKai,WuXiaojun,andYinHefeng(SchoolofInternetofThingsEngineering,JiangnanUniversity,Wuxi,Jiangsu214122)

Abstract Visionproblemrangingfromimageclusteringtomotionsegmentationcannaturallybeframedassubspacesegmentationproblem,inwhichoneaimstorecovermultiplelowdimensionalsubspacesfromnoisyandcorruptedinputdata.Lowrankrepresentation‐basedsubspacesegmentationalgorithm(LRR)formulatestheproblemasaconvexoptimizationandachievesimpressiveresults.However,itneedstotakealongtimetosolvetheconvexproblem,andtheclusteringaccuracyisnothighenough.Therefore,thispaperproposesadistributedlowrankrepresentation‐basedsparsesubspaceclusteringalgorithm(DLRRS).DLRRSadoptsthedistributedparallelcomputingtogetthecoefficientmatrix,thentaketheabsolutevalueofeachelementofthecoefficientmatrix,andretaintheklargestcoefficientspercolumnandsettheotherelementsto0togetanewcoefficientmatrix.Finally,DLRRSperformsspectralclusteringoverthenewcoefficientmatrix.Butitdoesn摧thaveincrementallearningfunction,sothereisascalabledistributedlowrankrepresentation‐basedsparsesubspaceclusteringalgorithm(SDLRRS)here.Ifnewsamplesarebroughtin,SDLRRScanusetheformerclusteringresulttoclassifythenewsamplestogetthefinalresult.ExperimentalresultsonARandExtendedYaleBdatasetsshowthattheimprovedalgorithmscannotonlyobviouslyreducetherunningtime,butalsoachievehigheraccuracy,whichverifiesthattheproposedalgorithmsareefficientandfeasible.Keywords lowrankrepresentation;subspaceclustering;parallelcomputing;incrementallearning;coefficientsreconstruction

摘 要 针对基于低秩表示的子空间分割算法运算时间较长、

聚类的准确率也不够高,提出一种基于分布式低秩表示的稀疏子空间聚类算法(distributedlowrankrepresentation‐basedsparsesubspaceclusteringalgorithm,DLRRS),该算法采用分布式并行计算来得到低秩表示的系数矩阵,然后保留系数矩阵每列的前k个绝对值最大系数,其他系数置为0,用此系数矩阵构造一个稀疏的样本关系更突出的相似度矩阵,接着用谱聚类得到聚类结果.但是其不具备增量学习功能,为此再提出一种基于分布式低秩表示的增量式稀疏子空间聚类算法(scalabledistributedlowrankrepresentationbasedsparsesubspaceclusteringalgorithm,SDLRRS),如果有新增样本,可以利用前面的聚类结果对新增样本进行

相关主题