基于人工神经网络的数字音频水印研究(IJEME-V2-N1-4)

I.J.EducationandManagementEngineering2012,1,23-28PublishedOnlineJanuary2012inMECS()DOI:10.5815/ijeme.2012.01.04Availableonlineat:digitalaudiowatermarking;artificialneuralnetworks;discretecosinetransform©2012PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheInternationalConferenceonE-BusinessSystemandEducationTechnology1.IntroductionWiththerapiddevelopofcomputernetworkandmultimediainformationprocessingtechnology,itismoreeasilytocopydigitalproductillegally,thusintellectualpropertyprotectionfordigitalproductsisrequiredmoreurgently.Digitalwatermarkingtechnologyprovidesamethodofcopyrightprotectionforitssecurity,robustnessandimperceptibility.Digitalaudiowatermarkingtechnologyusesthefeaturesofhumanauditorysystemtoembedthesecretinformationintotheaudiofilesandachievesthepurposeofcopyrightprotectionfordigitalaudioproducts.Accordingtothedifferentwaysofembeddingthewatermarkintotheaudiofiles,currentdigitalaudiowatermarkingtechnologiescanbedividedintotimedomainalgorithmsandtransformdomainalgorithms.Theformerdirectlyembedsthewatermarkinformationintotheselectedtimedomainofaudiosignal,suchastheleastsignificantbitalgorithm[1],thephasecodingalgorithm[2],andtheechohidingmethod[3];thelatterperformssomekindoftransformationontheaudiosignal,andthenembedsthewatermarkintothetransformedcoefficients.Finallyitrecoversthewatermarkedaudiosignalthroughthecorrespondinginversetransform.TransformationalgorithmsaremainlythediscreteFouriertransform(DFT),thediscretecosinetransform(DCT),thediscretewavelettransform(DWT).Ingeneralspeaking,thetransformdomainalgorithmhasmoretransparency,security,andcapacityinembeddingwatermarkintoaudiosignal.Boththetimedomainalgorithmsandthetransformdomainalgorithmsachievethepurposeofembeddingwatermarkinformationintoaudiosignalbymodifyingtheoriginalaudiosignal.AlthoughitmakesfulluseofthecharacteristicsofhumanauditoryCorrespondingauthor:E-mailaddress:[e-mail]24DigitalAudioWatermarkingBasedonArtificialNeuralNetworksmodelsystems,theperceptionqualityofaudiosignalischangedtoacertainextent.Thispapermakesfulluseofthelearningandadaptivecapabilitiesofartificialneuralnetworksandusestheimportantcharactersofaudiosignalastheinputvectorofartificialneuralnetworks.Therelationshipbetweenaudiosignalandwatermarkinformationisestablishedthroughthelearningofartificialneuralnetworks,whichachievesembeddingwatermarkintotheoriginalaudiosignalwithoutmodifyingtheaudiodata.Thismethoddoesnotchangetheperceptionqualityofaudiosignalandimprovetheimperceptibilityofwatermark.Inaddition,thisproposedwatermarkingmethoddoesnotrequiretheoriginalaudiosignalforwatermarkingextraction.2.BackpropagationalgorithmofartificialneuralnetworksBackpropagation(BP)algorithmofartificialneuralnetwork[4]belongstoδalgorithm,whichisasupervisedmachinelearningalgorithm.ThemainideaofBPalgorithmistopropagatetheoutputlayererrorfrombacktofrontandindirectlycalculatethehiddenlayererror.BPalgorithmisdividedintotwophases.Inthefirstphase,thevalueofeachunitofoutputlayerisobtainedbycalculatingtheinputvectorfrominputlayertooutputlayer.Inthesecondstage(backpropagation),usingthevectorofoutputlayer,theerrorofeachhiddenlayeriscalculatedwhichisusedtomodifytheconnectionweightsofneuralnetwork.BPalgorithmusuallyusesgradientmethodtomodifytheweightsofneuralnetworkthatminimizethesumofsquarederrors.Inaddition,backpropagationalgorithmoftenusesSigmoidfunctionasoutputfunction.Fig.1showstheconventionalsymbolsinbackpropagationalgorithm.Forthecalculatingunitj,thesubscriptiisonbehalfofcalculatingunitiofitsformerlayerandthesubscriptkisonbehalfofcalculatingunitkofitslaterlayer.TheOjrepresentstheoutputvalueofcurrentlayerandtheWijrepresentstheweightfromtheformerlayertocurrentlayer.Whensampledataisinputtedtotheneuralnetwork,eachcalculatingunitfrominputlayertooutputlayerperformsthefollowingcalculation:Figure1.theconventionalsymbolsinback-propagationalgorithmiiijjOWnet(1))(jjnetfo(2)Fortheoutputlayer,thejyˆ(jjoyˆ)istheactualoutputvalueandthejyistheidealoutputvalue.TheoutputerrorEisasfollow:2j)ˆy(21jjyE(3)Inordertosimplifytheformula,thegradientisdefinedasfollow:jjnetE(4)kiWijWjOO····DigitalAudioWatermarkingBasedonArtificialNeuralNetworks25ConsideringtheimpactofweightsWij