基于变分高斯混合模型的图像分割算法.pdf

《基于变分高斯混合模型的图像分割算法.pdf》由会员上传分享，免费在线阅读，更多相关内容在应用文档-天天文库。

1、宁波大学学报(理工版)首届中国高校优秀科技期刊奖第27卷第1期，2014年1月浙江省优秀科技期刊一等奖Vo1．27No．1，Jan．2014JOURNALOFNINGBOUNIVERSITY(NSEE)ImageSegmentationviaVariationalMixtureofGaussionsZHANGYuan—yuan，ZHONGYi—weirFacultyofInformationScienceandTechnology，NingboUniversity，Ningbo315211，China)Abstra

2、ct：Gaussianmixturemodel(GMM)hasbeenefectivelyusedinimagesegmentation·Inthiscase，thefeaturesofanimagearedescribedbyamixturemodelwithKdiferentcomponents．However,howtochOOsethenumberofmixturecomponentsKandestimatemodelparametersarestillshortofsolutions·CurrentaIg

3、orithmssuchasmaximumlikelihoodandsamplingmethodsareknownfortheirownlimitations·SowepresentanalternativealgorithmbasedonBayesianvariationalmethodandapplyitinimagesegmentation．Thismethodworksatlesscomputationalcostthansamplingmethods，andcanalsonaturallyhandlemem

4、odelselectionproblem．Inthemodel’siterativeprocess，thealgorithmcanautomaticallydetenninethenumberofmixturecomponentsinviewofthedatacollected．BycomparingourmethodagainstotherclassicalsegmentationmethodsonnaturalimagesacquiredfromBerkeleySegmentationDataSet，itsug

5、geststhatourmethodprovidesbetterperformanceonimagesegmentation·Kevwords：imagesegmentation；variationalinference；Gaussianmixturemodels；expectation—maximizationCLCnumber：TP391．41Documentcode：AArticleID：1001—5132(2014)01—0023—06Imagesegmentationisanimportantproble

6、minnentsinordertoavoidover-fittingorunder-fiting．Ifcomputervisionwhichisaimtopartitionanimageintothechoiceofnumberisinappropriate，thespeedofdisjointregions，afterthatthepixelsinthesameregionconvergencewillbeveryslowandtheresultofhavesimilarvisualcharacteristics

7、．Recently，Gaussianclusteringmaybepoor．Manyinformationrulesaremixturemodelissuchausefulmathematicalmode1usedselectedtotrytosolvethisproblem[6-10]suchas，forunsupervisedimagesegmentationJ．ThismethodBayesianInformationCriterion(BIC)，AkaikeInfo卜modelstheimagewithmi

8、xtureofGaussians，andthemationCriterion(AIC1whichcomputerelativelOSSsegmentationisaprocessofassigningeachpixeltotheinforlTlation．Butthevdonotknowwhatthetruemodelcomponentwhichitmost