简介:Inthispaperwepresentanonmonotonetrustregionmethodfornonlinearleastsquaresproblemswithzero-residualandproveitsconvergenceproperties.Theextensivenumericalresultsarereportedwhichshowthatthenonmonotonetrustregionmethodisgenerallysuperiortotheusualtrustregionmethod.
简介:Thispaperpresentsatrustregiontwo-phasemodelalgorithmforsolvingtheequalityandboundconstratinednonlinearoptimizationproblem.Aconceptofsubstationarypointisgiven.Undersutableassumptions.theglobalconvergenceofthisalgorithmisprovedwithoutassumingthelinearindependenceofthegradientofactiveconstraints.Anumericalexampleisalsopresented.
简介:Inthispaper,wepresentanefficientapproachforunsupervisedsegmentationofnaturalandtexturalimagesbasedontheextractionofimagefeaturesandafastactivecontoursegmentationmodel.Weaddresstheproblemoftextureswhereneitherthegray-levelinformationnortheboundaryinformationisadequateforobjectextraction.Thisisoftenthecaseofnaturalimagescomposedofbothhomogeneousandtexturedregions.Becausetheseimagescannotbeingeneraldirectlyprocessedbythegray-levelinformation,weproposeanewtexturedescriptorwhichintrinsicallydefinesthegeometryoftexturesusingsemi-localimageinformationandtoolsfromdifferentialgeometry.Then,weusethepopularKullback-Leiblerdistancetodesignanactivecontourmodelwhichdistinguishesthebackgroundandtexturesofinterest.Theexistenceofaminimizingsolutiontotheproposedsegmentationmodelisproven.Finally,atexturesegmentationalgorithmbasedontheSplit-Bregmanmethodisintroducedtoextractmeaningfulobjectsinafastway.Promisingsyntheticandreal-worldresultsforgray-scaleandcolorimagesarepresented.
简介:Aclassoftrustregionmethodstorsolvinglinearinequalityconstrainedproblemsispropo6edinthispaper.Itisshownthatthealgorithmisofglobalconvergence.Thealgorithmusesaversionofthetwo-sldedprojectionandthestrategyoftheunconstrainedtrustregionmethods.Itkeepsthegoodconvergencepropertiesoftheunconstrainedcaseandhasthemeritsoftheprojectionmethod.Insomesense,ouralgorithmcanberegardedasanextensionandimprovementoftheprojectedtypealgorithm.
简介:Thispaperproposesanonmonotonicbacktrackingtrustregionalgorithmviabilevellinearprogrammingforsolvingthegeneralmulticommodityminimalcostflowproblems.Usingthedualitytheoryofthelinearprogrammingandconvextheory,thegeneralizeddirectionalderivativeofthegeneralmulticommodityminimalcostflowproblemsisderived.Theglobalconvergenceandsuperlinearconvergencerateoftheproposedalgorithmareestablishedundersomemildconditions.