radartargetrecognitionmethodusingimprovedsupportvectormachines.pdf

RadarTargetRecognitionMethodUsingImprovedSupportVectorMachines BasedonPolarizedHRRPs*XiaoHuaitie NationalUniversityof DefenseTechnology, SchoolofElectronicScience andEngineering, Changsha,Hunan,China, 410073 htxiao@GuoLei NationalUniversityof DefenseTechnology, SchoolofElectronicScience andEngineering, Changsha,Hunan,China, 410073FuQiang NationalUniversityof DefenseTechnology, SchoolofElectronicScience andEngineering, Changsha,Hunan,China, 410073AbstractTargetrecognitionbasedonhighrangeresolution (HRR)polarizedradarusingsupportvectormachines (SVMs)wasstudiedinthispaper.Afuzzymembership functionwasconstructedbasedonSVMdecision- makingfunctioninordertoimprovetheperformance ofOAAandOAOclassifiersformulti-classtarget,and HRRradartargetrecognitionmethodusingimproved SVMwasproposed.First,thepolarizedradar backscatterechoeswereprocessedbyincoherent integrationandpower-normalized,thelocationand lengthoftargetinechoeswereestimatedandrange profilesoftargetwereinterpolatedtocertainradial length,thenpolarizedprofileswereintegrated consideringtherelevancyofrangeprofilesofsame targetindifferentpolarizationstate,atlast,the improvedOAAandOAOclassifierswereusedfor targetclassification.Simulationexperimentresults showthattheproposedmethodhastheadvantageoflittlecapacityofcomputationandcanimprovethe performanceofclassifierseffectively.1.IntroductionRadartargetrecognitionistoidentifytheunknown targetfromitsradarechoedsignatures.Withthe increasedavailabilityofHRRradar,therehasbeena renewedinterestinit.ForHRRradar,projectionof targetscatteringcenteralongradarradialrangeaxis canbeobtained,whichiscalledhighrangeresolution profile(HRRP)oftarget[l].HRRPcontainsmoredetailtargetgeometryshapeandphysicsstructure informationthanthatofthelowrangeresolutionradar echoes,targetcanbeclassifiedbyusingHRRP[2].For HRRPidentificationapplication,matched-filtering technologyisusuallyused,whichcorrelatedtheknown HRRPfilterstothereceivedunknowntargetHRRP. AsisknownthatradarHRRPisastrongfunctionof targetaspect,andtargetmayexistatanypositionin realsystem.Inrealprocess,manyfiltersareneededfor matchingHRRPofdifferenttargetaspect,anditbrings problemsoflargecomputationanddatamemoryspace Supportvectormachine(SVM)isanewgeneration learningsystemproposedbyVapnik[3].Thetheoryof SVMisbasedontheideaofstructuralrisk minimization(SRM)[3]principleofstatisticallearning theory.ASVMclassifierfirstmapstheinputvectors intoahighdimensionalfeaturespaceandfindsan optimalseparatinghyperplaneinthefeaturespacethat maximizesthemarginbetweentwoclassesinthis space.Ithasrevealedextraordinaryadvantagein solvingsmallamountofpatterns,nonlinearandhigh dimensionalpatternrecognitionproblems.Inmany applications,SVMhasbeenshowntoprovidehigher performancethantraditionallearningmachines[4]and hasbeenintroducedaspowerfultoolsforsolving classificationproblems.It'sattractivecharacteristic motivateustousethiskindofclassifierforradartarget recognitionbasedonHRRP. Generally,SVMwasoriginallydesignedforbinary classificationproblem.Themulti-classclassification problemiscommonlysolvedbydecompositiontoTheprojectissupportedbyNationalNaturalScienceFoundationofChina(No.60572138),KeylabFoundation(No.9140C8001010601)and 973Program(No.51314).1-4244-0605-6/06/$20.00C2006IEEE.702severalbinaryproblemsinwhichthestandardSVM algorithmcanbeused.TherearetwotypicalSVM algorithmwhicharecalledone-against-all(OAA)and one-against-one(OAO)respectively[5].Accordingto Vapnik'sformulation,inOAASVM,ak-class problemisconvertedintoktwo-classproblemsandfor theithtwo-classproblem,classiisseparatedfromthe remainingclasses.Butbythisformulation unclassifiableregionsexistifweusethediscrete decisionfunctions.Onewaytosolvethisproblemisto usecontinuousdecision,anotheristointroducefuzzy membershipfunctions[6][7]. TakuyaInoue[6]proposedtheconceptionoffuzzy supportvectormachineanddefinedtruncated polyhedralpyramidalmembershipfunctionsusingthe decisionfunctionsobtainedbytrainingtheSVMto resolveunclassifiableregions.DaisukeTsujinishi][7] discussedfuzzyleastsquaresSVMthatresolve unclassifiableregionsformulticlassproblemsand definedamembershipfunctionusingtheminimumor averageoperation. Inmanyapplications,someinputpointsmaynotbe exactlyassignedtooneofthesetwoclasses.Someare moreimportanttobefullyassignedtooneclasssothat SVMcanseparatethesepointsmorecorrectly.Some datapointscorruptedbynoisesarelessmeaningfuland themachineshouldbettertodiscardthem.SVMlacks thiskindofability.Chun-FuLin[8]applyafuzzy membershiptoeachinputpointofSVMand reformulateSVMintofuzzySVMsuchthatdifferent inputpointscanmakedifferentcontributionstothe learningofdecisionsurface. InHRRradartargetrecognitionapplication,the definitionoffuzzyfunctionproposedin[6-8]wastoo complicatedandnotconvenientforcalculation.For resolvingunclassifiableregionsinHRRradartarget recognition,weconstructedafuzzymembership functionbasedonSVMdecision-makingfunction,and modifiedtheOAAandOAO SVMalgorithmbyusing thefuzzymembership. Therestofthispaperisorganizedasfollows:A briefreviewofthetheoryofSV。