基于概率神经网络的不规则图案智能几何分类(IJIGSP-V7-N4-2)

I.J.Image,GraphicsandSignalProcessing,2015,4,19-27PublishedOnlineMarch2015inMECS()DOI:10.5815/ijigsp.2015.04.02Copyright©2015MECSI.J.Image,GraphicsandSignalProcessing,2015,4,19-27IntelligentGeometricClassificationofIrregularPatternsviaProbabilisticNeuralNetworkSogandHoshyarmaneshAmirkabirUniversityofTechnology/DepartmentofMechanicalEngineering,Tehran,158754413,IranE-mail:[e-mail]ooniSharifUniversityofTechnology/CenterofExcellenceinHydrodynamics&DynamicsofMarineVehicle,113658639,Tehran,IranE-mail:[e-mail]rUniversityofTechnology/DepartmentofMechanicalEngineering,Tehran,158754413,IranE-mail:[e-mail]—Thispaperdealswithinterpretationofpatternsvianeuralnetworksunderorganizationandclassificationapproaches.Fiftydifferentgroupsofimagesincludinggeometricshapes,mechanicalinstruments,machines,animals,fruits,andotherclassesofsamplesareclassifiedhereintwosuccessivesteps.Eachprimarycategoryisdividedintothreedifferentsub-groups.Thepurposeisidentifyingtheclassandsub-classofeachinputsample.Nowadays,industryandmanufacturingaremovingtowardsautomation;henceaccuratedescriptionofphotosresultsinamyriadofindustrial,security,andmedicalapplicationsandtakesapressingpartinartificialintelligence’sprogression.Intelligentinterpretationofstructure’sdesigninCNCmachineeventuatesinautonomousselectionofcuttingtoolsbywhichanystructurecaneasilybemanufactured.Anyhow,thispapercomesupwithapatterninterpretationmethodtobeappliedinsubmarinedetectionpurposes.Remotelyoperatedvehicles(ROV)areusedtodetectandsurveyoilpipelinesandunderwatermarinestructures,somentionedneuralnetworkclassificationisapracticabletoolfordetectionmechanismandavoidingobstaclesinROVs.IndexTerms—PatternRecognition;NeuralNetwork;FeatureExtraction;DistanceHistogram.I.INTRODUCTIONMulti-classpatternclassificationemphasizesissuesincludingneuralnetworkarchitecture,decipheringschemes,andtrainingmethodology[1]usingeitheracombinationofsomesmallneuralnetworksoralargeone[2].FirstessentialstepofobstacleavoidancecontrolinAUVsistodetectthembasedonvision.Inthismethodthereceivedimagefromcameraisremotelytransferredtothecontroller,andthenimagesareprocessedtorecognizetheobstacles.Theobstacledetectionisthemostcriticalstageofroboticcontrolaccomplishedviacollectingandanalyzingthereceiveddata.Aclassificationmethodemployingtwoconsecutivesteppedneuralnetworksisusedtocategorizethereceivedimagesbasedontheboundaryinterpretationofdifferentpatterns.Accuracyofclassnumbersisverifiedbytheavailabledatabase;ultimately,applyingthisprocessmakesunderwatervehiclescapableofunderwaterobstacles’interpretationandavoidingthem.AmongsttheANNsbeingtrainedtocarryoutfeatureextraction[3-6],SOMs[4,7],andHopfieldANNs[8]haveapparentlydrawnresearchers’attention.Feed-forwardANNshavealsoplayedasignificantroleinlotsofthereviewedpatternclassificationissues[9-12].Neuralnetworksapplicationsarewidespreadfortheiracceptableperformanceinclassificationandfunctionestimation.Theyhavebeenutilizedwithsuccessinmedicalimageanalysis[13]fordiagnosisofdiseases[14]heartssoundclassification[15],automaticfacerecognition[16]andothervariousapplications.Asuitabledeductionoffeatureextractionisthechiefkeyinimageprocessing’sefficiency.Aninnovativeneuro-fuzzynetworkissuggestedwhichemploysasimpleFouriertransformmethodtoextractfeatures[17].TheproperLearningrulesguaranteestheresults’accuracy,forinstance,Hebbianlearningruleisusedinachaoticneuralnetwork[18]andgivenanacceptableperformance.Manyresearcheshavebeendonetomodifytheneuralnetworkfunctionbypresentingnewaccuratealgorithms.TheSVMisrelativelyanewpatternclassificationapproachbasedontheideasoflargemarginandmappingdataintoahigherdimensionalspace,aspointedoutin[19,20].Elsewhere,aneuralnetworkmodelisproposedbasedontheconceptofmulti-layerperceptrontrainedunderlimitationoflayersweightstoacertainrangetomeettheleastsquarederror[21].ImageInterpolationalgorithmsaresimulatedandlearnedbyadjustingweightsandbiasvaluesofneuralnetworksbasedoncameraidentificationmethods[22].patternclassificationapproach[23,24]withdifferenttypesof20IntelligentGeometricClassificationofIrregularPatternsviaProbabilisticNeuralNetworkCopyright©2015MECSI.J.Image,GraphicsandSignalProcessing,2015,4,19-27neuralnetworkhasalreadybeenderived;somehigher-orderneuralnetworksfordistortioninvariantpatternrecognition[25],identifyingsalientfeaturesforclassificationinbothfeedforwardneuralnetworks[26]andprobabilisticneuralnetworks[27],classificationofrelevantfeaturesintheinputvectorbyapplyingradialbasisneuralnetwork[28],andemployingmultilayerperceptronneuralnetworkfortheclassificationofremote-sensingimages[29,30]areseveralcitablestudiesinthiscase.TheoperationalregionsofROVsincludingoilpipelines,underwatertankersandreservoirsofoilandfreshwateronseabottomandweldedbondingofmarinestructurescanbedetectedandsurveyedbytherobotsprogrammedwithimageclassificationsystem,sotheoperationiscompletelyheldintelligentwithoutanyuserinterference.Imagescouldbecapturedonlinebyacameramoun