Two-stage Character Classification: a Combined Approach of Clustering and Support Vector Classifiers

A COMBINED APPROACH OF CLUSTERING AND SUPPORT VECTOR CLASSIFIERS LOUIS VUURPIJL AND LAMBERT SCHOMAKER vuurpijl ni i.kun.nl,s homaker ni i.kun.nl, http://hwr.ni i.kun.nl This paper des ribes a two-stage lassi ation method for (1) lassi ation of isolated hara ters and (2) veri ation of the lassi ation result. Chara ter prototypes are generated using hierar hi al lustering. For those prototypes known to sometimes produ e wrong lassi ation results, a \support ve tor lassi er" (sv ) is trained. The sv an be used to in rease the on den e that a lassi ation is orre t and furthermore de ide on a lassi ation if the on den e using the standard method is too low. Experiments with the iUF UNIPEN database yield 94% re ognition rate. In ases where both lassi ers agree, the error rate is zero. 1 Introdu tion In handwriting re ognition, a standard approa h of hara ter lassi ation is to (1) nd a number of hara ter prototypes (allographs), based on a set of distin tive features and (2) mat h unknown hara ters to the labeled allographs for lassi ation. A wide range of te hniques an be used to generate a set of prototypes like lustering methods, nearest-neighbor methods or neural networks. Su h methods have the advantage over, e.g., hidden-Markov models or multi-layered per eptrons, in that the prototypes are visible: they are made expli it and the lassi ation performan e of ea h individual prototype an be investigated in detail. Clustering is a well-known te hnique for nding a set of prototypes. In1, a novel lustering te hnique was des ribed, whi h obtains a set of prototypes, organized in a hierar hi al stru ture. Ea h prototype ontains members of similar hara ter shapes, and its entroid is de ned as the average of its members. Using prototypes for distan e-based nearestentroid mat hing, lassi ation performan es of 86% were a hieved, where the lassi ation errors were partly due to labeling errors and mainly due to the onfusion that ertain ombinations of lasses introdu e. The aspe t of \zooming in" on onfused hara ter lasses is further elaborated in2. In that paper, a system is des ribed whi h engages soalled intelligent agents in ase of potential onfusion. Here, we des ribe a method of ombining the luster-based prototype mat hing approa h with a lass-separation te hnique alled Support Ve tor Classi ation3;4;5. The method is used to (1) in rease the on den e that a lassi ation is orre t and (2) de ide on a lassi ation if the on den e using the nearestentroid mat hing is too low. 423