圖像檢索中基于SVM的相關(guān)反饋技術(shù)研究.pdf

1、西北工業(yè)大學(xué)碩士學(xué)位論文圖像檢索中基于SVM的相關(guān)反饋技術(shù)研究姓名：胡珊申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：模式識(shí)別與智能系統(tǒng)指導(dǎo)教師：郭雷20070301西北工業(yè)大學(xué)碩士論文AbstractAbstractThedevelopmentofmultimediaandintemettechniqueshasbroughtonarapidincreaseofthenumberofdigitalimagesThetraditionaltextbasedi

2、mageretrievalfacessomeproblemswhendealingwithlargescaleimagedatabasesItisnecessarytoextractmorecomprehensive，moregeneralandmoreobjectiveimagefeaturestoconductimageretrievalContentbasedimageretrieval(CBIR)wasproposedtomee

3、tsucharequirementIt’SatechniquethatretrievesrelevantimagesbasedonimagecontentsuchascolortextureandshapefeaturesThispaperfocusesonSVMbasedrelevancefeedbackinCBIRThemainworkofthispaperissunlmarizedasfollows：1TheSVM—basedre

4、levancefeedbackteehniquesarestudiedandcomparedwiththereweightingmethodusedintheMARSsystemTogetaroundthesmallsampleissuethefeedbackexamplesa托accumulatedaftereachiterationAnewfeedbackmethodusingmulti—SVMisproposedtoimprove

5、retrievalperformanceTheexperimentalresultsshowthattheSVMbasedtechniqueshavewonderfulperformanceinimageretrievalusingthefeaturesandimagedatabasechosenbythispaper2AnewHSVcolorspacequantizationmethodisproposedInthismethodwh

6、enintensityvalueorsaturationvalueissmallerthanacertainthreshold，colorisapproximatedbyagrayvalue；otherwisecolorisrepresentedbyitshueandsaturationthustheinfluenceofintensityisignoredThisquantizationmethodreducesthedimensio

7、nofcolorhistogramandexperimentshaveshownitsgoodperformanceininlageretrieval3Multi—featurebasedimageretrievalisconductedColortextureandshapefeaturesareextractedfromimagesandcombinedtetrieverelevantimagesExperimentsshowtha

8、tusingmulti—featureachievesmuchbetterretrievalperformancethanusingasingleimagefeature4AcontentbasedimageretrievalsystemisdesignedThissystemallowsusertoselecttheimagefeaturesbeforeconductingtheretrievalUsersCanchoosecolor

9、texture，orshape，oranycombinationofthesethreefeaturesThesystemalsosupportsrelevancefeedbackTworelevancefeedbackmethodsarcavailable：SVMbasedandre—weightingKeywords：contentbasedimageretrieval，multi—feature，relevancefeedback