图像语义特征的提取与分析

上海交通大学硕士学位论文图像语义特征的提取与分析姓名：张好申请学位级别：硕士专业：信号与信息处理指导教师：李生红20061201ICBIRCBIRCBIR/HSV+Luv+Tamura/Lab+Luv/II82.17%ABSTRACTIIIExtractionandAnalysisofSemanticImageFeaturesABSTRACTWiththedevelopmentofnetworkandmultimediatechnologies,moreandmoreimagedatabaseshavebeencomeforth,moreover,thesizeoftheseimagedatabaseshasbecomelargerthanbefore.Moreuserswanttoretrieveimagestheyneedfromthehugeimagedatabase.Basedonthisbackground,atechnology,namedContent-BasedImageRetrieval(CBIR),hasbeenfastandwelldeveloped.However,ontheonehand,thetraditionalCBIRsystemsdon’tconsiderthesemanticinformationofimages.Ontheotherhand,wecan’tuseallthefeaturesduetothehighdimensionandcomplexcomputation.So,analysisandselectionofsemanticimagefeaturesisnecessary,afterfeatureextractioninordertochoosethehigh-discriminativefeatures,whichwillbeusefultoimprovetheretrievalaccuracy.ThisthesisdoestheresearchontheextractionandanalysisofsemanticimagefeaturesandproposesanewmethodtoanalyzethediscriminativeskillsoffeaturesusingMutualInformation,providingbasisonsemanticimagefeatureselection.ThismethodhasasolidtheoryfoundationandanalyzesfeaturesABSTRACTIVaccordingtothefeature’sinformation.Itisalsoindependentonthetypesofclassifiersanddistributionoflabels.Therefore,thismethodbasedonMutualInformationhasastrongpopularityandpracticability.Firstly,thisthesisintroducescurrentresearchtrends,CBIRsystemarchitecturesandkeytechnicalknowledge.Theprimarysemanticimageclassificationtechnologiesandextractionmethodsofimagefeaturesarealsomentioned.Then,weextractfeaturesincludingcolor,textureandedgeforlandscape/human,indoor/outdoor,building/landscapeclassificationproblems.AndintensiveanalysisofthediscriminativeskillsaboutfeaturesbasedonMIvaluesisgivenwiththeexperimentalresultsonlargeimagedatabases.Furthermore,weselectthemost-discriminatingfeaturesintoaset.Thenwedotheexperimentusingclassifiererroranddistancemeasurementtovalidatethefeasibilityofthenewmethodusingmutualinformation.Thenewmethodweproposedovercomesthedefectsofothermethods.Theapplicationofthemost-discriminatingfeaturesbasedonourconclusionusedtoclassifyakindofhierarchicalimagedatabaseindicatestheimportanceandinfluenceoftheextractionandanalysisofsemanticimagefeatures,theaccuracyofwhichreached82.17%.Keywords:featureanalysis,mutualinformation,semanticfeature,imagefeature20071162007116200711611.12005119CNNIC[1]CBIR[2]1-11-1Fig.1-1ModelofImageSemanticLevel21.21.2.170[3,4,5,6]90CBIR,3[7]1-21-2Fig.1-2ArchitectureofContent-BasedImageRetrievalSystem1-1CBIR1.2.21.2.2.1QBIC[8,9]IBMAlmaden/4QBICRGBYIQLabMTMkTamuraCoarsenessDirectionalityContrastQBICQBICKLTR*VisualSEEKWebSEEK[10,11]VisualSEEKquad-treeR-treeWebSEEKWeb//WebSEEKVirage[12,]VIRAGEQBIC010VirageOracleSybaseRetrievalWare[13]ExcaliburExcalibur5605Photobook[14]FourEyesNetra[15]UCSBGaborBlobworld[16]R*MARS[17]UIUCMARS1.2.2.2CBIRVailaya[18]90.8%695.3%Chapelle[19]Coral386501200625300358300HSV16161616%RBFSVMAekasandraMojsiovicBemiceRogowitz[20]Gorkani[21]9844Amoid[22]Li[23]Szummer[24]K-N-N132490%7Biederman[25]Schyns[26]1-31-3Fig.1-3Relationshipamongimage,low-levelimagefeaturesandsemanticfeatures1.38632003AA1421608910SVM2.12.1.12.1.1.1RGBRGBHSVHSVHueSaturationValueRGBHSV11[27](2-1)Max=max(R,G,B)Min=min(R,G,B)V=0.299R+0.587G+0.114B0,Max=0Max-Min/Max,S=0,Max=Min60(G-B)/(Max-Min),Max=RGBH=360+60(G-B)/(Max-Min),Max=RGB60(2+(B-R)/(Max-Min)),Max=G60(4+(R-G)/(Max-Min)),(2-1)S,V[0,1],H[0,360]CIELABCIELUVRGBCIELAB[28]2-2=BGRZYX990.0010.0000.0011.0813.0177.0200.0310.0490.0−=otherwise'3.903008856.0'if16)'(1163/1*YYYL(2-2)()3/123/11*500KKa−=()3/133/12*200KKb−=otherwise008856.0if116/16787.7Φ+ΦΦ=iiiiKi=1,2,3Φ1=X’=X/X0,Φ2=Y’=Y/Y0,Φ3=Z’=Z/Z0RGBCIELUV[28]LLuvL)'(130**uuLu−=12)'(130**vvLv−=(2-3))315(4'ZYXXuu++==,'1.56/(153)vvvYXYZ==++2.1.1.2HSVnHh1,h2,…,hnhjj2.1.1.3[29]njjj.(j,j)(1,j),,(n,n)132.1.1.4ColorMoments[30]meanvarianceskewness11221133111(())1(())NiijjNiijijNiijijpNpNSpNµσµµ=====−=−∑∑∑(2-4)ijpji2.1.1.5[31,32,33,34]bin142.1.2TamuraGabor2.1.2.1TamuraTamura[35]CoarsenessContrastDirectionalityLinelikenessRegularityRoughnessTamura1Tamura1)2k2k11112121222(,)(,)/2kkkkyxkkixjyAxygij−−−−+−+−=−=−=∑∑(2-5)k=1,2,3,4,5211,11,(,)(2,)(2,)(,)(,2)(,2)kkkhkkkkkvkkExyAxyAxyExyAxyAxy−−−−=+−−=+−−(2-6),(,)khExy,(,)kvExyk15(,)2kbestSxy=(2-7)3bestS111(,)mncrsbestijFSijmn===×∑∑(2-8)2Tamura444µασ=(2-9)144conFσα=(2-10)4µ2σconF3Tamura1)()/2GHV∆=∆+∆(2-11)arctan(/)2VHπθ=∆∆+(2-12)H∆V∆3×3101101101−−−111000111−−−2)G∆0.53)2162.1.2.2[36,37]Pi,j|d,dijPi,j|d,2,{(,|,)}ijASMPijdθ=∑(2-13),{(,)|,)}log{(,|,)}ijENTPijdPijdθθ=−•∑(2-14)ENTENT∑−=jidjiPjiCON,2),|,()(θ(2-15)CONCONyxjiyxdjiPjiCORσσθµµ∑−−=,),|,())(((2-16)xµyµxσyσPxPyPy),|,(θdjiPPx),|,(θdjiP172.1.2.3GaborGaborManjunathMa[38]GaborPWTTWTMARSARGabor222211(,)()exp(()2)22xyxyxygxyjwxππσσσσ=−++(2-17)Gaborg(x,y)(,)(','),1,,,'(cossin),'(sincos),/mmnmmgxyagxyamnxaxyyaxynkθθθθθπ−−−==+=−+=(2-18)kma−m2.1.2.4simultaneousauto-regressiveSARMarkovMRF[39]SARsg(s)()sε()()()()rDgsrgsrsµθε∈=+++∑(2-19)µDs()rθ()sε0182σθσσsθSAR2.1.32