基于颜色布局描述子、灰度共生矩阵和-k-近邻的内容图像检索(IJITCS-V12-N3-3)

I.J.InformationTechnologyandComputerScience,2020,3,19-25PublishedOnlineJune2020inMECS()DOI:10.5815/ijitcs.2020.03.03Copyright©2020MECSI.J.InformationTechnologyandComputerScience,2020,3,19-25Content-BasedImageRetrievalUsingColorLayoutDescriptor,Gray-LevelCo-OccurrenceMatrixandK-NearestNeighborsMd.FarhanSadiqueComputerScienceandEngineeringDiscipline,KhulnaUniversity,Khulna,BangladeshE-mail:[e-mail]ScienceandEngineeringDiscipline,KhulnaUniversity,Khulna,BangladeshE-mail:[e-mail]:07December2019;Accepted:25December2019;Published:08June2020Abstract—Content-basedimageretrieval(CBIR)istheprocessofretrievingsimilarimagesofaqueryimagefromasourceofimagesbasedontheimagecontents.Inthispaper,colorandtexturefeaturesareusedtorepresentimagecontents.Colorlayoutdescriptor(CLD)andgray-levelco-occurrencematrix(GLCM)areusedascolorandtexturefeaturesrespectively.CLDandGLCMareefficientforrepresentingimageswithlocaldominantregions.Forretrievingsimilarimagesofaqueryimage,thefeaturesofthequeryimageismatchedwiththatoftheimagesofthesource.Weusecityblockdistanceforthisfeaturematchingpurpose.K-nearestimagesusingcityblockdistancearethesimilarimagesofaqueryimage.OurCBIRapproachisscaleinvariantasCLDisscaleinvariant.Anothersetoffeatures,GLCMdefinescolorpatterns.Itmakesthesystemefficientforretrievingsimilarimagesbasedonspatialrelationshipsbetweencolors.Wealsomeasuretheefficiencyofourapproachusingk-nearestneighborsalgorithm.Performanceofourproposedmethod,intermsofprecisionandrecall,ispromisingandbetter,comparedtosomerecentrelatedworks.IndexTerms—Colorlayoutdescriptor,gray-levelco-occurrencematrix,KNN,corel-1k,dominantcolor,scaleinvariant.I.INTRODUCTIONThenecessityofmakingcomputerhuman-likeisincreasinginday-to-daylifewiththeincreasedamountofcomputerapplications.Makingcomputerthatperceivesthingslikehumandoesisthemainconcernofcomputervision.Content-basedimageretrieval(CBIR)isoneoftheimportantareasofcomputervision.Ingeneral,CBIRmeansretrievingsimilarimagesofaqueryimageaccordingtothecontentsoftheimage.Italsomeansrecognizingimagesaccordingtotheircontents.TheinvestigationofCBIRhasbeengoingonsince1990s[10].ToshikazuKato[11]proposedasimilargraphicsymbolretrievalsystemofagivensketch.Thissystemusesspatialdistributionofthegraylevel,spatialfrequency,localcorrelationmeasureandlocalcontrastmeasureforrepresentingthecontentsofthesketchandgraphicsymbols.MariaTzelepiandAnastasiosTefas[10]proposedadeepconvolutionalneuralnetworkbasedCBIRsystem.DeepconvolutionalneuralnetworkbasedCBIRsystemsareconsideredasoneofthemosteffectiveoptionsforbuildingaCBIRsystem.However,tobuildaneffectiveCBIRsystemusingdeepneuralnetwork,alotoftrainingdataarerequired.Moreover,thereisnowaytoknowwhyasetoffeaturesareselectedaftertraining.In[15],FamaoYeetal.proposedaremotesensingimageretrievalsystemwhichusesconvolutionalneuralnetwork(CNN)forclassificationpurpose,whereasmostoftheworksuseCNNforextractingfeatures.S.M.MohidulIslamandRameswarDebnath[12]proposedarotation,scaleandtranslationinvariantCBIRapproach.Thisapproachusescolormomentsandwaveletentropyasimagefeatures.InourproposedCBIRmethod,wehaveusedcolorlayoutdescriptorandgray-levelco-occurrencematrixforrepresentingthecontentsoftheimages.Scaleinvariantcolorlayoutdescriptorrepresentscolorandgray-levelco-occurrencematrixrepresentstextureofanimage.Wehaveusedk-nearestneighbors(KNN)algorithmforsimilarimageretrievalandfortrainingandtestingimagesforperformanceevaluation.Ourproposedmethodoutperformssomerelatedmethodsintermsofprecision,recall,F-scoreandaccuracy.Moreover,webelievethatourfeatureextractionmethodsarebetterforthefollowingreasons:Colorlayoutdescriptorextractscolorinformationfromseveralblocksofanimage.Thus,itprovideslocalcolorinformation.Therefore,localdominantcolormakesalargeimpactinthisfeature.Itissuitableforrepresentingimagescontainingfewobjectsinfrontofalargebackground.20Content-BasedImageRetrievalUsingColorLayoutDescriptor,Gray-LevelCo-OccurrenceMatrixandK-NearestNeighborsCopyright©2020MECSI.J.InformationTechnologyandComputerScience,2020,3,19-25Colorlayoutdescriptorisscaleinvariant.Gray-levelco-occurrencematrixgeneratesfeaturesbasedonthespatialrelationshipbetweenpixels.Itissimplebutveryusefulwhenanimagecontainsoneormorenotableareaswithsamecolorpattern.Therestofthepaperisorganizedasfollows.SectionIIdescribessomerelatedworks.SectionIIIoutlinesourproposedmethodindetails.ExperimentalresultsandcomparisonareshowninsectionIV.SectionVconcludesourwork.II.RELATEDWORKSSawetSomnugpongandKanokwanKhiewwan[7]proposedaCBIRsystemusingcolorcorrelogramsandedgedirectionhistogram.Colorcorrelogramsrepresenttheinformationabouttherelativelocationsofcolorvalues.Edgedirectionhistogrammakesthesystemsuitableformatchingsimilarimagesofdifferentcolors.Fivetypesofshapesaredeterminedusingedgedirectionhistogram.Euclideandistanceisusedtofindthedistancebetweenthefeaturesoftheimages.ItusesCorel-1Kdataset[4,5]foritsperformanceevaluation.Thismethodshowsitsefficiencyusingprecisionandrecallvalues