简介:TheNeighborhoodPreservingEmbedding(NPE)algorithmisrecentlyproposedasanewdimensionalityreductionmethod.However,itisconfinedtolineartransformsinthedataspace.Forthis,basedontheNPEalgorithm,anewnonlineardimensionalityreductionmethodisproposed,whichcanpreservethelocalstructuresofthedatainthefeaturespace.First,combinedwiththeMercerkernel,thesolutiontotheweightmatrixinthefeaturespaceisgottenandthenthecorrespondingeigenvalueproblemoftheKernelNPE(KNPE)methodisdeduced.Finally,theKNPEalgorithmisresolvedthroughatransformedoptimizationproblemandQRdecomposition.Theexperimentalresultsonthreereal-worlddatasetsshowthatthenewmethodisbetterthanNPE,KernelPCA(KPCA)andKernelLDA(KLDA)inperformance.
简介:Amajordifficultyinmultivariablecontroldesignisthecross-couplingbetweeninputsandoutputswhichobscurestheeffectsofaspecificcontrollerontheoverallbehaviorofthesystem.Thispaperconsiderstheapplicationofkernelmethodindecouplingmultivariableoutputfeedbackcontrollers.Simulationresultsarepresentedtoshowthefeasibilityftheproposedtechnique.
简介:Previously,anovelclassifiercalledKernel-basedNonlinearDiscriminator(KND)wasproposedtodiscriminateapatternclassfromotherclassesbyminimizingmeaneffectofthelatter.Toconsidertheeffectofthetargetclass,thispaperintroducesanobliqueprojectionalgorithmtodeterminethecoefficientsofaKNDsothatitisextendedtoanewversioncalledextendedKND(eKND).IneKNDconstruction,thedesiredoutputvectorofthetargetclassisobliquelyprojectedontotherelevantsubspacealongthesubspacerelatedtootherclasses.Inaddition,asimpletechniqueisproposedtocalculatetheassociatedobliqueprojectionoperator.ExperimentalresultsonhandwrittendigitrecognitionshowthatthealgorithmperformesbetterthanaKNDclassifierandsomeothercommonlyusedclassifiers.
简介:Withthevigorousexpansionofnonlinearadaptivefilteringwithreal-valuedkernelfunctions,itscounterpartcomplexkerneladaptivefilteringalgorithmswerealsosequentiallyproposedtosolvethecomplex-valuednonlinearproblemsarisinginalmostallreal-worldapplications.ThispaperfirstlypresentstwoschemesofthecomplexGaussiankernel-basedadaptivefilteringalgorithmstoillustratetheirrespectivecharacteristics.ThenthetheoreticalconvergencebehaviorofthecomplexGaussiankernelleastmeansquare(LMS)algorithmisstudiedbyusingthefixeddictionarystrategy.ThesimulationresultsdemonstratethatthetheoreticalcurvespredictedbythederivedanalyticalmodelsconsistentlycoincidewiththeMonteCarlosimulationresultsinbothtransientandsteady-statestagesfortwointroducedcomplexGaussiankernelLMSalgonthmsusingnon-circularcomplexdata.Theanalyticalmodelsareabletoberegardasatheoreticaltoolevaluatingabilityandallowtocomparewithmeansquareerror(MSE)performanceamongofcomplexkernelLMS(KLMS)methodsaccordingtothespecifiedkernelbandwidthandthelengthofdictionary.
简介:Inpractice,retrainingatrainedclassifierisnecessarywhennoveldatabecomeavailable.ThispaperadoptsanincrementallearningproceduretoadaptivelytrainaKernel-basedNonlinearRepresentor(KNR),arecentlypresentednonlinearclassifierforoptimalpatternrepresentation,sothatitsgeneralizationabilitymaybeevaluatedintime-variantsituationandasparserrepresentationisobtainedforcomputationallyintensivetasks.Theaddressedtechniquesareappliedtohandwrittendigitclassificationtoillustratethefeasibilityforpatternrecognition.
简介:Undertheframeworkofsupportvectormachines,thispaperproposesanewkernelmethodbasedonneighborbandsmutualinformationforhyperspectraldatumclassification.Thisalgorithmassignsweightstodifferentbandsinthekernelfunctionaccordingtotheamountofusefulinformationthattheycontain,whichmakesthebandwithmoreusefulinforma-tionplaymoreimportantroleintheclassification.Ourresearchhasshownthatthebandwithgreatermutualinformationbetweenneighborbandscontainsmoreusefulinformation,andhenceweusethemutualinformationofeachbandanditsneighborbandsastheweightsoftheproposedkernelmethod.Theexperimentalresultsshowthatforthesupportvectormachinesbasedonpolynomialandradialbasisfunction,afterintroducingtheproposedkernelfunction,theaverageaccuracyisincreasedmorethan1.2%withoutusinganyreferencemaporincreasingmuchmorecomputationaltime.
简介:Inthispaper,weproposeanewmethodthatcombinescollageerrorinfractaldomainandHumomentinvariantsforimageretrievalwithastatisticalmethod-variablebandwidthKernelDensityEstimation(KDE).TheproposedmethodiscalledCHK(KDEofCollageerrorandHumoment)anditistestedontheVistextexturedatabasewith640naturalimages.ExperimentalresultsshowthattheAverageRetrievalRate(ARR)canreachinto78.18%,whichdemonstratesthattheproposedmethodperformsbetterthantheonewithparametersrespectivelyaswellasthecommonlyusedhistogrammethodbothonretrievalrateandretrievaltime.