Manifold Transfer Subspace Learning (MTSL) for Applications in Aided Target Recognition

Abstract

Thisarticledescribeshowtransfersubspacelearninghasrecentlygainedpopularity foritsabilitytoperformcross-datasetandcross-domainobjectrecognition.Theability toleverageexistingdatawithouttheneedforadditionaldatacollectionsisattractive formonitoringandsurveillancetechnology,specificallyforaidedtargetrecognition applications. Transfer subspace learning enables the incorporation of sparse and dynamicallycollecteddataintoexistingsystemsthatutilizelargedatabases.Manifold learninghasalsogainedpopularityforitssuccessatdimensionalityreduction.Inthis contribution,Manifoldlearningandtransfersubspacelearningarecombinedtocreate anewsystemcapableofachievinghightargetrecognitionrates.Themanifoldlearning technique used in this contribution is diffusion maps, a nonlinear dimensionality reductiontechniquebasedonaheatdiffusionanalogy.Thetransfersubspacelearning techniqueusedisTransferFisher’sLinearDiscriminativeAnalysis.Thenewsystem, manifold transfer subspace learning, sequentially integrates manifold learning and transfersubspacelearning.Inthisarticle,theabilityofthenewtechniquestoachieve high target recognition rates for cross-dataset and cross-domain applications is illustratedusingavarietyofdiversedatasets. KeywoRdS Diffusion Maps, Manifold Learning, Target Recognition, Transfer Learning, Transfer Subspace Learning