ADAPTATION OF PROTOTYPE SETS IN ON-LINE RECOGNITION OF ISOLATED HANDWRITTEN LATIN CHARACTERS
Jorma T. Laaksonen, V. Vuori, Erkki Oja, Jari A. Kangas · Series in machine perception and artificial intelligence · 1999
ADAPTATIONOFPROTOTYPESETSINON-LINERECOGNITIONOFISOLATEDHANDWRITTENLATINCHARACTERSJormaLAAKSONEN,VuokkoVUORI,andErkkiOJAHelsinkiUniversityofTechnologyLaboratoryofComputerandInformationScienceP.O.BOX5400,FIN-02015HUT,FinlandE-mail:fjorma.laaksonen,vuokko.vuori,erkki.ojag@hut. JariKANGASNokiaResearchCenterP.O.Box100FIN-33721Tampere,FinlandE-mail:jari.kangas@research.nokia.comResultsonacomparisonofadaptiverecognitiontechniquesforon-lineofhandwrittenLatinalphab etsarepresented.Theemphasisison veadaptivclassi cationstrategiesdescrib edinthispap er.Thearebasedon rstgeneratingauser-indep endentsetofprototyp echaractersandthenmo difyingthissetinordertoadaptiteachuser'sp ersonalwritingstyle.Theinitialisformedbyasimpleclusteringalgorithm.Themo di cationoftheprototyp esetisp erformedusingthreemo desofop eration:1)newprototyp esareadded,2)exist-ingprototyp esarereshap edtob ettermatchtheinput,and3)whicpro ducefalseclassi cationsareremoved.Theclassi cationdecisionusesthek-NearestNeighb or(k-NN)ruleforthedistancesb etweenunknowncharacterandthestoredprototyp es.Thedistancesarecalculatedbyusingtemplatematch-ingwithDynamicTimeWarping(DTW).Thereshapingoftheexistingprototyp esisp erformedbyutilizingamo di edversionoftheLearningVectorQuantization(LVQ)algorithm.Thepresentedexp erimentsshowthattherecognitionsystemisabletoadaptwelltheuser'swritingstylewithonlyafew{sayonehundredhandwrittencharacters.1Intro ductionWepresenttheresultsofaseriesexp erimentswhichassesseddevel-opmentofanadaptiveclassi erforon-linerecognitionhandwrittenLatincharacters.Theaimofthestudyhasb eentoevaluatep otentialato-tallyunsupervisedadaptivehandwritingrecognizer.Thesystemshouldb eabletoutilizetheuserinputenhanceclassi cationaccuracysimultaneouslywhenitisused.Noseparatetrainingp erio dshouldthereforeb enecessary.Thesystemwoulduseitsownrecognitionsinaself-supervisedfashion:therecognitionswhicharenotindicatedasfailuresbytheuserregardedcorrect.Similarly,theuser-rep ortedmisclassi cationsareusedtorevise1