基于互信息和神经网络识别图(无花果L.)(IJIGSP-V5-N9-8)

I.J.Image,GraphicsandSignalProcessing,2013,9,50-57PublishedOnlineJuly2013inMECS()DOI:10.5815/ijigsp.2013.09.08Copyright©2013MECSI.J.Image,GraphicsandSignalProcessing,2013,9,50-57Fig(FicusCaricaL.)IdentificationBasedonMutualInformationandNeuralNetworksGhadaKattmahGeneralCommissionofAgriculturalScientificResearch,DepartmentofHorticulturescience,Syriae-mail:[e-mail]ofComputers&Informatics,CanalSuezUniversity,Egypte-mail:[e-mail]—Theprocessofrecognitionandidentificationofplantspeciesisverytime-consumingasithasbeenmainlycarriedoutbybotanists.Thefocusofcomputerizedlivingplant'sidentificationisonstablefeature’sextractionofplants.Leaf-basedfeaturesarepreferredoverfruits,alsothelongperiodofitsexistencethanfruits.Inthispreliminarystudy,westudyandproposeneuralnetworksandMutualinformationforidentificationoftwo,threeFigcultivars(FicusCaricaL.)inSyriaregion.TheidentificationdependsonimagefeaturesofFigtreeleaves.AfeatureextractorisdesignedbasedonMutualInformationcomputation.TheNeuralNetworksisusedwithtwohiddenlayersandoneoutputlayerwith3nodesthatcorrespondtovarieties(classes)ofFIGleaves.Theproposaltechniqueisatesteronadatabaseof84imagesleaveswith28imagesforeachvariety(class).Theresultshowsthatourtechniqueispromising,wheretherecognitionrates100%,and92%forthetrainingandtestingrespectivelyforthetwocultivarswith100%and90forthethreecultivars.Thepreliminaryresultsobtainedindicatedthetechnicalfeasibilityoftheproposedmethod,whichwillbeappliedformorethan80varietiesexistentinSyria.IndexTerms—Patternrecognition,Textureanalysis,neuralnetwork,mutualinformation,FigtreeclassificationandidentificationI.INTRODUCTIONThecomputerizationofplantmanagementspeciesbecomesmorepopular.Theprocessofrecognitionandidentificationofplantspeciesisverytime-consumingasithasbeenmainlycarriedoutbybotanists.Thefocusofcomputerizedlivingplantidentificationisonstablefeature’sextractionofplants.Theinformationofleafplaysanimportantroleinidentifyinglivingplants.Leaf-basedfeaturesarepreferredoverowner’sfruits,etc.duetotheseasonalnatureofthelaterandalsotheabundanceofleaves.Fig(FicusCaricaL.)treeplaysanimportantroleintheSyrianeconomy,whereSyriaproduces42944tonsayear.Forthisreason,theclassificationoffigtreespeciesisofparamountimportancetopreservetypeandimproveproduction.Inthisresearch,wedescribedtheclassificationofaplantFig(FicusCaricaL)withmultiplevarieties.FigisoneoftheoldestfruittreesintheMediterraneanzone.Itswildgeneticresourcestillexistsinmanycountries,includingSyriaandAnatolia,whichareconsideredasthenaturalhabitatsofthefigtree.Theclassificationofplantleafhasbeenintroducedinseveralapproachessuchask-NearestNeighborClassifier(k-NN),ProbabilisticNeuralNetwork(PNN),GeneticAlgorithm(GA),SupportVectorMachine(SVM),andPrincipalComponentAnalysis(PCA)[1].Zulkifli[2]proposedGeneralRegressionNeuralNetworktoclassify10kindsofplantswithgreencolorleaves.Wuetal.[3]usedPNNtoclassify32kindsofgreenleaves.Singhetal.[4]suggestedSVMtoimplementaclassifierfortheproblem.Shabanzadeetal.[5]usedLinearDiscriminantAnalysis(LDA).Severalofresearchersareusedaspectratio,leafvein,andinvariantmomenttoidentifyplant.Severalfeaturessuchasanaspectratio(ratiobetweenlengthandwidthofleaf),theratioofperimetertothediameterofleaf,andveinfeatureswereusedtocharacterizetheleafwithaccuracyof90%.Wuetal.[3,6]Ontheotherhand,Textureanalysisisimportantinmanyapplicationsofcomputerimageanalysisforclassification,detection,orsegmentationofimagesbasedonlocalspatialvariationsoftheintensityorthecolor.Actually,shape,colorandtexturefeaturesarecommonfeaturesinvolvedinseveralapplications,suchasin[7]and[8].However,someresearchersusedpartofthosefeaturesonly.InvariantmomentsproposedbyHueareverypopularinimageprocessingtorecognizeobjects,includingleavesofplants[9-10].PolarFouriertransforms(PFT)proposedbyZhang[2].Severalleafclassificationsystemshaveincorporatedtexturefeaturestoimprovetheperformance,suchasin[12]thatusedentropy;homogeneityandcontractionderivedfromtheco-occurrencematrixcamefromDigitalWaveletTransform(DWT),in[13]thatusedlacunaritytocapturethetextureofleafandin[14]thatusedGLCM.Texturefeaturescanbeextractedbyusingvariousmethods.Gray-levelGaborFilterandLocalbinaryFig(FicusCaricaL.)IdentificationBasedonMutualInformationandNeuralNetworks51Copyright©2013MECSI.J.Image,GraphicsandSignalProcessing,2013,9,50-57pattern(LBP)areexamplesofpopularmethodstoextracttexturefeatures.Toimplementthepreliminaryclassificationsystem,inthisresearch,wetriedtocapturethetextureoftheleafbyanewfeatureextractiontechniquebasedonmutualinformation.Thosefeatureswereinputtedintotheidentificationsystemthatusesaneuralnetworkclassifier.Testingwasdonebyusingadataset.Theresultshowsthatmethodimprovesgoodperformance(92%)oftheidentificationsystemwith2cultivarsand90with3cultivars.MutualInformation(MI)andEntropyprovideanintuitivetooltomeasuretheuncertaintyofrandomvariablesandtheinformationsharedbythem,inwhichtheentropyandthemutualinformationaretwocriticalconcepts.Theentropyisameasureoftheuncertaint