Randomized Distribution Feature for Image Classification

《Randomized Distribution Feature for Image Classification》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、426ECAI2016G.A.Kaminkaetal.(Eds.)©2016TheAuthorsandIOSPress.ThisarticleispublishedonlinewithOpenAccessbyIOSPressanddistributedunderthetermsoftheCreativeCommonsAttributionNon-CommercialLicense4.0(CCBY-NC4.0).doi:10.3233/978-1-61499-672-9-426RandomizedDistribut

2、ionFeatureforImageClassificationHongmingShanandJunpingZhang∗,1Abstract.andthemetric-basedone.Thehistogram-basedmodelusuallyrep-Localimagefeaturescanbeassumedtobedrawnfromanun-resentseachimagebytheempirical,one-dimensionalhistogramthatknowndistribution.Forimage

3、classification,suchfeaturesarecom-enumeratestheoccurrenceprobabilityofeachpointsetinthebagparedthroughthehistogram-basedmodelorthemetric-basedmodel.ofvisualwords.Here,thecollectionofthesewordsiscalledacode-Byquantizingtheselocalfeaturesintoasetofhistograms,the

4、bookordictionary.Thedisadvantagesofthismethodarethatthehistogram-basedmodelisconvenientandhasvectorialrepresenta-sizeofcodebookisdifficulttoselect,andthecomputationalcostoftionofimagebutinformationcouldbelostinvectorquantization.generatingthecodebookbythequant

5、izationalgorithmsisexpensive.Unlikethehistogram-basedmodel,themetric-basedmodelestimatesBesides,theinformationwillbelostinthequantizationprocess[34].themetricsovertheunderlyingdistributionoflocalfeaturesimmedi-Incontrast,themetric-basedmodelestimatesstatistic

6、almetricsoverately,achievingbetterpredictiveperformance.However,themodeltheunderlyingdistributionofimageswithhigheraccuracy.Thead-requireshighercomputationalcostandlosesthebenefitofvectorialvantageofthismodelisthatitdoesnotrequirequantizationtech-representatio

7、nofimage.niquesandselectingthesizeofcodebook,eachofwhichcouldresultToretaintheadvantagesofthesetwomodels,thispaperproposesinthelossofperformanceinimageclassification.However,thesethe(doubly)randomizeddistributionfeaturesthatrepresenttheun-metricssufferfromhigh

8、computationalcostsincetheyoperateoverderlyingdistributionoflocalfeaturesineachimageasavectorialpairwisesamples.AnotherdrawbackofthemodelisthatthematricesfeaturebyutilizingrandomFourierfea