局部特征描述子算法研究

上海交通大学硕士学位论文局部特征描述子算法研究姓名：施鹏申请学位级别：硕士专业：模式识别与智能系统指导教师：姚莉秀20080101III123IVVSTUDYONLOCALDESCRIPTORAbstractAlongwiththefastgrowthofthewaysandamountofhumans’imagegaining,somebasicresearchesindigitalimageprocessingbecamemoreandmoreimportant.Amongthem,localdescriptorisoneofthemostbasicandmostdifficultfieldsincomputervisionandpatternrecognition.Generally,localfeatureisanimagepatternwhichdiffersfromitsimmediateneighborhood.Theyaredistinctive,donotrequiresegmentationandrobusttoocclusion,overlap,geometrictransformationandilluminationchange.Becauseofthesecharacteristics,localfeaturebecomethemostimportanttechnologyinfeatureextraction.Localfeatureshaveproventobeverysuccessfulinapplicationssuchaswidebaselinematching,imageretrieval,objectrecognition,texturerecognition,robotlocalization,videodatamining,imagemosaicandrecognitionofobjectcategories.Throughdecades’researchinginthisfieldthetechnologyoflocalfeaturehasmadegreatprogress,theseapproachesfirstdetectfeaturesandthencomputeasetofdescriptorsforthesefeatures,afterthat,wecantransformtheimagematchingproblemintofeaturevectormeasurement.However,therestillexistssomedefectsinthelocalfeatureextraction,suchashigh-dimensionalvectors,ignoringtheglobalinformation.Therefore,localfeaturetechniquesarenowfarfromsophisticated.Inthispaper,aconclusionaboutpreviousachievementsinlocalfeaturesismade.Theresearchofthispapermainlyfocusesonthedetectionoffeaturepointsandtheconstructionoffeaturedescription,includingashortreviewofthesetechniques.Meanwhile,anovellocaldescriptorispresented.Additionally,weprovideaninstanceofrealapplicationoflocalfeaturetechnologyinimageindex.Themaincontributionofthispaperare:1.Afteracarefulanalysisaboutthepopularfeaturedetectionalgorithms,wepresentanovelmethodbasedonsalientdetection,thisimprovementcanenhancetheaccuracyoftheinterestpointslocalityandmakethelocalfeaturereflecttheessenceinformationoftheimagemoreprecisely.VI2.Asolutionthatisbasedoncolornaming,intensityvalueandintensitygradientispresentedfortheconstructionofthelocalfeature.Thisnovellocaldescriptionisinvarianttoilluminationandgeometrytransformation,itcanalsorecognizetheobjectswhicharesimilarintextureandintensityvaluewhiledifferentincolor.3.Atwo-stageimageindexstrategybasedoncolorhistogramandlocaldescriptorispresented.Thismethodcanreducethealgorithmcomplexity.Keywords:Imageindex,localdescriptor,colorhistogram,salientdetectionI2009225II??v“v”2009225200922511.1[1,2][3,4][5,6][7][8][9][10]1.21.2.11.2.21954Attneave[11]21.2.3[12,13,14]GregMori[15][16,17,18,19][20,21]JohnsonHebert[22][23]ZabihWoodfill[24]DavidLowe(sift)[6]3Sift[25]Geometrichistogram[26]shapecontextBeaudetHessian[27],hessian4.1.2[28][29][30][31]KrysitanHarris-Laplace[33]Laplace1.31.3.1490(contentbasedimageretrieveCBIR)[34][35]1.3.2SwainBallardFuntFinlaysonHealeySlaterIBMQBICMITPhotobookVisualSEEKSiggelkowBurkhardt1.3.35tinneJosef/K-means,meanshift,binning1.4siftsurfHarris-Affine672.12.2()()(Pr)HimageHInnovationHiorKnowledge=+()Hinnovation(Pr)HiorKnowledge()LIlog()I()AI8()RI()PI()SI()AI()*()LIhI1111()1113111hI=()RI()()LIAI-2()[exp(()()]SIiFFTRIPI=+Figure.2.1Detectingresultsindifferentscales.Figure(a)istheinputimage,figure(b)istheresultsin64*64scale,figure(b)istheresultsin128*128scale,figure(b)istheresultsin64*649scale.2.2Figure2.2Salientdetectionresult,theleftistheinputimage,therightisthesalientregion.()sx()Ox1()()0ifSxthresholdOxotherwise⎧=⎨⎩(2-5)(())*3thresholdEsx=(2-6)(())Esx2.3(a)inputimage(b)salientmap(c)objectmap10(d)(e)(f)(d)inputimage(e)salientmap(f)objectmap(g)(h)(i)(g)inputimage(h)salientmap(i)objectmap2.3Figure.2.3Salientdetectionresultsbasedonexperienceformula2.32.4(a)25(b)50(a)(b)11(c)(d)2.4Figure.2.4Grayhistogramofthesalientmap25612561([]*)graygraygraygrayhistgraygraythresholdgray=====∑∑(2-7)gray[]histgray01256102.442.12.1(a)2510026(b)508237(c)508940(d)4070292.1Table.2.1Comparisonofexperiencethresholdandweightedthreshold122.5(a)(b)(c)(a)inputimage(b)experiencethresholdresults(c)weightedthresholdresults(e)(f)(d)inputimage(e)experiencethresholdresults(f)weightedthresholdresults(g)(h)(i)(g)inputimage(h)experiencethresholdresults(i)weightedthresholdresults2.5Figure.2.5comparisonofthetwoalgorithms2.4133.1Koenderink[38]22212()exp22xygssps+-=3-1(,,)()*(,)LxygIxyss=3-23-2x,y*I(x,y)Lindeberg(difference-of-Gaussian)[39]14(,,)((,,)(,,))*(,)(,,)(,,)DxyGxykGxyIxyLxykLxysssss=-=-3-3Lowe3-1Figure.3.1Constructthescalespace(1)2(2)k0skkss=g3-1k0.5152SS33-1kk,1/2Sk=DOG16×16k3DOG3-1(3-2)DOG3-1DOGDOGDOG89263-2263-2Figure.3.2Localizationoftheinterestpoints3.2[40]3.2.1HarrisHarrisharrisstephens[40]16222(,)(,)(,)()*(,)(,)(,)xDxDyDDIxDyDyDIxIxIxMgIxIxIxssssssss⎡⎤=⎢⎥⎢⎥⎣⎦3-4(,)()*()xDDxIxgIxss∂=∂3-522212()exp22xygssps+-=3-6det()()CornernessMtraceMl=-(3-7)12det()*Mll=12()*traceMll=dettracel0.04harris3.3harrisFigure.3.3Harriscornerdetectors173.2.2SusanSusanSmithBrady[41]SusanSmallestUnivalueSegmentAssimilatingNucleusUSANsusan3.4susanFigure.3.4Susancornerdetectors3.2.3harris-laplaceharris-laplace[42]harrislindeberg[43,44]3.5Laplacian[45]183.510.13.9Figure.3.5Example