Content Based Image Retrieval using Color and Text

Signal&ImageProcessing:AnInternationalJournal(SIPIJ)Vol.3,No.1,February2012DOI:10.5121/sipij.2012.310439ContentBasedImageRetrievalusingColorandTextureManimalaSingha*andK.Hemachandran$Dept.ofComputerScience,AssamUniversity,SilcharIndia.Pincode788011*[e-mail],$[e-mail]eedofcontentbasedimageretrievaltechniquecanbefoundinanumberofdifferentdomainssuchasDataMining,Education,MedicalImaging,CrimePrevention,Weatherforecasting,RemoteSensingandManagementofEarthResources.Thispaperpresentsthecontentbasedimageretrieval,usingfeaturesliketextureandcolor,calledWBCHIR(WaveletBasedColorHistogramImageRetrieval).Thetextureandcolorfeaturesareextractedthroughwavelettransformationandcolorhistogramandthecombinationofthesefeaturesisrobusttoscalingandtranslationofobjectsinanimage.TheproposedsystemhasdemonstratedapromisingandfasterretrievalmethodonaWANGimagedatabasecontaining1000general-purposecolorimages.Theperformancehasbeenevaluatedbycomparingwiththeexistingsystemsintheliterature.KeywordsImageRetrieval,ColorHistogram,ColorSpaces,Quantization,SimilarityMatching,HaarWavelet,PrecisionandRecall.1.INTRODUCTIONResearchoncontent-basedimageretrievalhasgainedtremendousmomentumduringthelastdecade.AlotofresearchworkhasbeencarriedoutonImageRetrievalbymanyresearchers,expandinginbothdepthandbreadth[1]-[5].ThetermContentBasedImageRetrieval(CBIR)seemstohaveoriginatedwiththeworkofKato[6]fortheautomaticretrievaloftheimagesfromadatabase,basedonthecolorandshapepresent.Sincethen,thetermhaswidelybeenusedtodescribetheprocessofretrievingdesiredimagesfromalargecollectionofdatabase,onthebasisofsyntacticalimagefeatures(color,textureandshape).Thetechniques,toolsandalgorithmsthatareused,originatefromthefields,suchasstatistics,patternrecognition,signalprocessing,dataminingandcomputervision.Inthepastdecade,manyimageretrievalsystemshavebeensuccessfullydeveloped,suchastheIBMQBICSystem[7],developedattheIBMAlmadenResearchCenter,theVIRAGESystem[8],developedbytheVirageIncorporation,thePhotobookSystem[9],developedbytheMITMediaLab,theVisualSeekSystem[10],developedatColumbiaUniversity,theWBIISSystem[11]developedatStanfordUniversity,andtheBlobworldSystem[12],developedatU.C.BerkeleyandSIMPLIcitySystem[13].Sincesimplycolor,textureandshapefeaturescannotsufficientlyrepresentimagesemantics,semantic-basedimageretrievalisstillanopenproblem.CBIRisthemostimportantandeffectiveimageretrievalmethodandwidelystudiedinbothacademiaandindustryarena.Inthispaperweproposeanimageretrievalsystem,calledWavelet-BasedColorHistogramImageRetrieval(WBCHIR),Signal&ImageProcessing:AnInternationalJournal(SIPIJ)Vol.3,No.1,February201240basedonthecombinationofcolorandtexturefeatures.Thecolorhistogramforcolorfeatureandwaveletrepresentationfortextureandlocationinformationofanimage.Thisreducestheprocessingtimeforretrievalofanimagewithmorepromisingrepresentatives.Theextractionofcolorfeaturesfromdigitalimagesdependsonanunderstandingofthetheoryofcolorandtherepresentationofcolorindigitalimages.Colorspacesareanimportantcomponentforrelatingcolortoitsrepresentationindigitalform.Thetransformationsbetweendifferentcolorspacesandthequantizationofcolorinformationareprimarydeterminantsofagivenfeatureextractionmethod.Colorisusuallyrepresentedbycolorhistogram,colorcorrelogram,colorcoherencevectorandcolormoment,undercertainacolorspace[14-17].Thecolorhistogramfeaturehasbeenusedbymanyresearchersforimageretrieval[18and19].Acolorhistogramisavector,whereeachelementrepresentsthenumberofpixelsfallinginabin,ofanimage[20].Thecolorhistogramhasbeenusedasoneofthefeatureextractionattributeswiththeadvantagelikerobustnesswithrespecttogeometricchangesoftheobjectsintheimage.Howeverthecolorhistogrammayfailwhenthetexturefeatureisdominantinanimage[21].LiandLee[22]haveproposedaringbasedfuzzyhistogramfeaturetoovercomethelimitationofconventionalcolorhistogram.Thedistanceformulausedbymanyresearchers,forimageretrieval,includeHistogramEuclideanDistance,HistogramIntersectionDistance,HistogramManhattanDistanceandHistogramQuadraticDistance[23-27].Textureisalsoconsideredasoneofthefeatureextractionattributesbymanyresearchers[28-31].Althoughthereisnoformaldefinitionfortexture,intuitivelythisdescriptorprovidesmeasuresofthepropertiessuchassmoothness,coarseness,andregularity.Mainlythetexturefeaturesofanimageareanalyzedthroughstatistical,structuralandspectralmethods[32].Therestofthepaperisorganizedasfollows:Insection2,abriefreviewoftherelatedworkispresented.Thesection3describesthecolorfeatureextraction.Thesection4,presentsthetexturefeatureextractionandthesection5,presentsthesimilaritymatching.Theproposedmethodisgiveninsection6andsection7describestheperformanceevaluationoftheproposedmethod.Finallytheexperimentalworkandtheconclusionsarepresentedinsection8andsection9respectively.2.RELATEDWORKLinetal.[14]proposedacolor-textureandcolor-histogrambasedimageretrievalsystem(CTCHIR).Theyproposed(1)threeimagefeatures,basedoncolor,textureandcolordistribution,ascolorco-occurrencematrix(CCM),differencebetweenpixelsofsca