欢迎来到天天文库
浏览记录
ID:37818091
大小:696.22 KB
页数:10页
时间:2019-05-31
《基于分级mean+shift的图像分割算法》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、万方数据计算机研究与发展ISSN1000—12391CN11.1777/TPJournalofComputerResearchandDevelopment46(9):1424—1431,2009基于分级meanshift的图像分割算法汤杨h2潘志庚1汤敏3王平安4夏德深21(浙江大学CAD&CG国家重点实验室杭州310027)2(南京理工大学计算机科学与技术学院南京210094)3(河海大学计算机及信息工程学院南京210098)‘(香港中文大学计算机科学与工程学系香港)(tang.yang@163.com)ImageSegm
2、entationwithHierarchicalMeanShiftTangYan91~,PanZhigen91,TangMin3,PhengAnnHen94,andXiaDeshenz1(StateKeyLaboratoryofCAD&CG。ZhejiangUniversity,Hangzhou310027)2(SchoolofComputerScience&Technology。NanjingUniversityofScience&Technology,Nanjing210094)3(CollegeofComputer&I
3、nformationEngineering,HohaiUniversity,Nanjing210098)‘(DepartmentofComputerScience&Engineering。TheChineseUniversityofHongKong,HongKong)AbstractConnectedchannelswerefoundsomewhereinthefeaturespaceintraditionalmeanshiftmethod.Consequently,toaccomplishthesegmentationby
4、onlyonetrialoftenleadstounsatisfactoryresult,especiallyinweakedgesorregionswithoutGaussiancharacter.Tosolvethisproblem,theauthorsdesignflhierarchicalsegmentmethodbasedonmeanshifttechnique.Uponontheoriginalsamplingdata,clusteringisperformediterativelywithdifferentba
5、ndwidthsonthecentersgainedpreviously.Thus。atreestructureisestablishedbetweenthenodesindifferentlevels.Accordingtotheinheritancerelationship,theleafnodesaremergedandclassifiedintodifferentcategoriesfinally.Themethodwasimplementedandtestedinbothgrayandcolorimages.Com
6、paredwiththetraditionalmeanshiftmethod,thehierarchicalmethodhasadvantagetoreserve、thedetaislinthesamescale.Also,althoughthemultipletimeclusteringisneeded,thecomputationcostwillnotincreaseobviouslyduetothedecreasingsamplesizesandvariedbandwidths.Thehierarchicalmeans
7、hiftmethodhasbeenprovedtobepromisingintheexperiments.Butmoretheoreticalanalysisisrequiredandthemethodalsoneedstobeimprovedbymoreexperiments.·Keywordsmeanshift;hierarchical;imagesegment;connectedchannel摘要实验发现传统meanshift算法进行分割时常会产生连接通道问题,使得几个分类簇之间无法完全分开.针对该问题,提出一种改进的
8、分级meanshift图像分割算法,在初次迭代获得的聚类中心基础上采用不同的带宽矩阵进行多次聚类,从而获得不同级的聚类中心集合,并建立一个归属树结构,最终通过叶节点与根节点的归属关系进行归类从而完成图像分割.实验证明改进算法可以更好地保留图像的局部信息,同时具有较好的适用性.关键词mean
此文档下载收益归作者所有