硕士论文-基于多特征信息融合的图像检索技术研究

华侨大学硕士学位论文基于多特征信息融合的图像检索技术研究姓名：叶宇光申请学位级别：硕士专业：计算机应用技术指导教师：陈锻生200609151CBIR90RGBHSIGaborlogSVMSVM2ABSTRACTContentBasedImageRetrieval(CBIR)hasbeenaveryactiveresearchareasince1990’s.Inthecontentbasedimageretrievalitisretrievedbysinglevisualcharacteratthebeginningoftheresearchsuchas:colortextureandfigurecharacters.Colorandtexturecharacteraretwoveryimportantandwidelyusedimagecharactersinthecontentbasedimageretrieval.Sofirstlythesetwofeaturesarerespectivelydiscussedinthispaper.Andatthesametimeinvariabletheoryhasachievedsuccessfulapplicationinthecontentbasedimagedataretrievalinrecentyearssoinvariantcharacterofimageretrievalisalsodiscussedinthispaper.Firsttwocolorspaces(RGBandHSV)threecolorhsitogramsmethods(traditionhsitogramaccumulationhsitogramlocalaccumulationhsitogram)andcolormatchingalgorithmarediscussedinthispaper.Thentexturefeatureisanalyzedanddiscussedandgaborwavelettransformationisusedtopickuptexturefeature.AfterwardsthispaperbrieflydescribesrelatedcontentofinvarianttheoryandchoosestheinvariantdescriptorofcombiningwithDFTandlogpolartransformationtobeimagecharacterdescriptoranddoestherelevantimageretrievalexperiment.Relevancefeedbacktechniquehasbeenanimportantapproachinimageretrievaldiscussedinthispaper.Meanwhilethispaperquotesanovelimageretrievalmethodofsupportvectormachinebasedrelevancefeedback,andtheproposedapproachpartiallysolvesthesampleinsufficiencyproblemthatexistsinSVMbasedrelevancefeedbackalgorithm,andshowtheapproachhasgoodperformanceforretrieval.Theimageretrievalfusingwithmulti-characterisaprogressdirection.Howeverhowtoefficientlyfuseandaffirmtheweightvalueamongdifferentcharactersinthetraditionalrelevancefeedbackisabigproblem.Thispaper,withtheideaofinformationfusion,proposesatwo-layerSVMclassifiermodeltoimprovetheprecisionoftheclassificationandusesSVM-basedInformationFusionMachinetoefficientlysolvetheproblemoffusingmulti-characterintheimageretrieval.Itradicallysolvesthebigproblemofdistillingandfusingmulti-character.Manyexperimentsprovethatthethreeimagecharacters,thatis,invariantcharacter,colorandtexturecharacter,usethefusionmethodwell.Furthermore,theeffectofimageretrievalimprovesalotcomparingwiththatofthetraditionalmethod.Keywords:ImageRetrievalInformationFusionRelevancefeedbackSupportVectorMachine2006.9.152006.12.202006.12.2011.1Internet()Content-basedImageRetrievalCBIR)1992[1]CBIRCBIR[2]Computervision(Imageprocessing)(Imageunderstanding)(Database)1.2902[3][4]1.2.1QBICIBMQBIC(QueryByImageContent)QBICQBIC(RGB)(YIQ)(LAB)MTM(Munsell)kTamuraQBICKLTR-QBIC()VirageQBICJegrey((Primitive))()()VirageEngine3PhotobookPhotobookMITPhotobookFourEyesPicard4NetraNetraUCSBNetraGaborMARSMARS(MultimediaAnalysisandRetrievalSystem)UIUC(UniversityofIllinoisatUrbanaChampaign)MARSMARSCORE1.2.2MiresMiresMiresSVMMires5ImageSeekImageSeekVQ(VectorQuantization)1.3CBIRCBIR(1)(2)(3)(1)2)(3)(4)(5)6(6)()1.4SVMSVMSVMSVM1Gabor2SVMSVMSVMSVMSVM3SVMSVM72.1[5]OnLineOffLine82-12.2MPEG-72.2.11)[6]Swain[7]HafnerL2[8]Stricker[9]Stricker[10]SmithChang2)()9MRSAR2090MaGabor[11][12]3)[13][14]4):1-D2-D[15]5)6)2102.2.2(relevancefeedback)2.2.311MinkowskiHausdorffXPRpnRxxxX⊂=},...,,{21Xxxlk∈,kldlkxx,d+→×RXXd:10≥kld20=kldlkxx=3lkkldd=4jlkjklddd+≤1.Minkowskyd(XY)=λλ11−∑=niiiyx)1(∝≤λ2-12.Manhattand(XY)=∑=−niiiyx12-2Minkowskyλ=13.Cityblockd(XY)=iniiiyxω∑=−12-3Manhattaniω)10(iω4.Euclideand(XY)=2112−∑=niiiyx2-4Minkowskyλ=2125.MahalanobisD2=()()YXCYXT−−−12-5XYC9A1BACCI6.HausdorffA{a1…ap}B{b1…bq}HausdorffHAB=maxh(AB)h(BA)2-6||||minmaxB)h(A,jiBbAabaji−=∈∈2-7||||minmaxA)h(B,ijAaBbabij−=∈∈2-82-6HausdorffHausdorffh(AB)ABHausdorffABh(BA)h(AB)Hausdorff1950FredAttneave[16]Tversky[17]1977AmosTversky(contrastmodel)[18]13TverskyabABTversky(contrastmodel)s3SfAB0abcdABCDS(ab)S(cd)s(ab)≥s(cd)S(ab)=f(A∩B)-αf(A-B)-βf(B-A)fABTverskyf2.2.4[19]20(TheCurseofDimension)[19]KLKarhunenLoveRVA-File314HashB-2.32.3.11231RecallRatePrecision234567TauABab15ccaaBAPAPABPrecall+=∪==)()()|(2-9baaBAPBPBAPpresise+=∪==)()()|(2-10RxPy-PVRPVRf(xy)f(xy)x-y∫=10),(dxyxfSfSfPVR[20]()PVR-[21](1)30.20.50.8[22](2)1111(3)0.5(4)BermanShapiro25400[23]efficiencyofretrievalNnT16≤=TNTnTNNnT//η2-11NTNT2-11NT2.3.231232.3.3123/4abcd2.41718rotationinvariancescale-invariance3.1MPEG-7colorspaceISO/IEC2001RGBYCbCrHSVHMMD[24]RGBHSI3.1.1RGBRGBRGBRGBRedGreenBlue24truecolor198(1)243RGB28+8+8=224=16777216RGBCRTRGBRGB{01255}RGB3-1RGBRGB3-1RGBRGBR3RGB3.1.2HSIHSIHue20SaturationIntensityMunseuHSIIHSHSI3-23-2HSIRGBHSI3[01]RGBHSIHSIRGBHSI3)],,[min(31))(()(2)()(arccos2))(()(2)()(arccos22bgribgrbgrsgbbgbrgrbrgrgbbgbrgrbrgrh++=++−=−−+−−+−−≤−−+−−+−=π3-1H21=+−=+−=−=),,max(34),,max(32),,max(bgrbsgrbgrgsrbbgrrsbghππ3-2R=G=BHSIIHSIHSI[25]HSIHRGBHSIH1-D[26]3.21991SwainBallardIL(C1C2…CL)CIiCIhih1h2…hLH(h1h2…hL)L1L23.2.1H22NnkHk=)(k=01L-1(3-3)kLknkN3-3abcdabcd3-33.2.21-D23NnkIiki∑==0)(k=01L-1(3-4)k=L-11133.2.3HSIH=0H=/3H=2/3H60H6[60k60k+1]k=01…53-13-160[17][30+60k30+60