搜索
I.J.Image,GraphicsandSignalProcessing,2014,5,28-35PublishedOnlineApril2014inMECS()DOI:10.5815/ijigsp.2014.05.04Copyright©2014MECSI.J.Image,GraphicsandSignalProcessing,2014,5,28-35High-speedImagecompressionbasedontheCombinationofModifiedSelf-organizingMapsandBack-PropagationNeuralNetworksOmidNaliDepartmentofElectricalEngineering,Saghezbranch,IslamicAzaduniversity,Saghez,IranE-mail:omid.nali@gmail.comAbstract—Thispaperpresentsahighspeedimagecompressionbasedonthecombinationofmodifiedself-organizingmapsandBack-Propagationneuralnetworks.Intheself-organizingmodelnumberoftheneuronsareinaflattopology.Theseneuronsininteractionformedself-organizingneuralnetwork.Thetaskthisneuralnetworkisestimatedadistributefunction.Finallynetworkdispersescellsintheinputspaceuntilestimatedprobabilitydensityofinputs.Distributeofneuronsininputspaceprobabilityisaninformationcompression.SointheproposedmethodfirstbyModifiedSelf-OrganizingFeatureMaps(MSOFM)weachieveddistributedfunctionoftheinputimagebyaweightvectortheninthenextstagetheseinformationcompressedareappliedtoback-propagationalgorithmuntilimageagaincompressed.Theperformanceoftheproposedmethodhasbeenevaluatedusingsomestandardimages.TheresultsdemonstratethattheproposedmethodhasHigh-speedoverotherexistingworks.IndexTerms—Imagecompression;Back-Propagationalgorithm;ModifiedSelf-OrganizingFeatureMaps.I.INTRODUCTIONThemainadvantageofcompressionisthatitreducesthedatastoragerequirements.Italsooffersanattractiveapproachtoreducethecommunicationcostintransmittinghighvolumesofdataoverlong-haullinksviahighereffectiveutilizationoftheavailablebandwidthinthedatalinks.Thissignificantlyaidsinreducingthecostofcommunicationduetothedataratereduction.Becauseofthedataratereduction,datacompressionalsoincreasesthequalityofmultimediapresentationsthroughlimited-bandwidthcommunicationchannels[1].Imagecompressionisoneoftheimportantchallengesintheinformationcompressionontheotherhandartificialneuralnetworkshavebecomepopularoverthelasttenyearsfordiverseapplicationsforimagecompressiontomachinevisionandimageprocessing.Highperformanceimagecompressionalgorithmsmaybedevelopedandimplementedinthoseneuralnetworks.Theneuralnetworkimagecompressionalgorithmcanbesummarizedasfollows:1-Back-Proagationimagecompression2-Hebbianlearningbasedimagecompression3-VectorQuantization(VQ)neuralnetworks.4-Predictiveneuralnetworks.Inthispapertheproposedmethodisbasedontwomajorstages,thefirststageisMSOFMandthesecondstageisback-propagationalgorithm.InthefirststagebyMSOFMandtheproposedweightupdatemethodestimatedadistributefunctionoftheinputimageinlowerdimensionsbyneurons(weightmatrix)thisisfirstdatacompression.Insecondstageweightmatrixisappliedtoback-propagationalgorithmuntilanotherstageofcompressionisappliedtotheinputimage.Figure1showsablockdiagramoftheproposedmethod.MSOFMNetworkBack-PropagationNetworkInputImageReconstructedImageEstimatedImageOnNeuronsFigure1.Blockdiagramoftheproposedmethod.AsseeninfigureoutputoftheMSOFMnetworkisestimatedimageonneuronsbasedontheinputimage.Sonumberofcolumnsarereducedthanthenumberofcolumnsoftheoriginalimage,theoutputofthisstageisasinputofback-propagationstage.Infollowingmoredetailoftheproposedmethodispresented.Theremainderofthepaperisorganizedasfollows:SectionIIfocusesonrelatedworksSectionIIIemphasizesontheproposedmethodandalsocomparison.ExperimentalresultsoftheproposedmethodarepresentedinsectionIV.FinallysectionVprovidestheconclusionofthispaper.II.RELATEDWORKSInrecentyearsmanymethodsforimagecompressionsystemshavebeenproposed.TheNeuralNetwork(NN)efficiencyforimagecompressionvariousworkswereHigh-speedImagecompressionbasedontheCombinationofModifiedSelf-organizingMapsand29Back-PropagationNeuralNetworksCopyright©2014MECSI.J.Image,GraphicsandSignalProcessing,2014,5,28-35suggestedinthepast.In[2]expoundstheprincipleofback-propagationneuralnetworkwithapplicationstoimagecompressionandtheneuralnetworkmodels.Thenanimagecompressionalgorithmbasedonanimprovedback-propagationnetworkisdeveloped.In[3]proposedanapproachformappingthepixelsbyestimatingtheCumulativeDistributionFunctionisasimplemethodofpre-processinganytypeofimage.Duetotheuniformfrequencyofoccurrenceofgraylevelsbythisoptimalcontraststretching,theconvergenceoftheback-propagationneuralnetworkisaugmented.In[4]presentsanovelcombiningtechniqueforimagecompressionbasedontheHierarchicalFiniteStateVectorQuantization(HFSVQ).In[5]presentsacompressionschemefordigitalstillimages,byusingtheKohonen’sneuralnetworkalgorithm,notonlyforitsvectorquantizationfeature,butalsoforitstopologicalproperty.In[6]afuzzyoptimaldesignbasedonneuralnetworksispresentedasanewmethodofimageprocessing.Thecombinationsystemadoptsanewfuzzyneuronnetwork(FNN)whichcanappropriatelyadjustinputandoutputvalues,andincreaserobustness,stabilityandworkingspeedofthenetworkbyachievingahighcompressionratio.In[7]presentedamethodthatisbasedonthecapabilitiesofmodifiedself-organizingKohonenneuralnetwork.In[8]approachofmappingthepixelsbyestimatingtheCumulativeDistributionFunctionispresentedthismethodi