Support Vector Machines for the Classification of Western Handwritten Capitals

OF WESTERN HANDWRITTEN CAPITALS FUSI WANG, LOUIS VUURPIJL AND LAMBERT SCHOMAKER Nijmegen Institute for Cognition and Information P.O.Box 9104, 6500 HE Nijmegen, The Netherlands E-mail: fswang(vuurpijl,s homaker) ni i.kun.nl http://hwr.ni i.kun.nl In this paper, new te hniques are presented using Support Ve tor Ma hines (SVMs) for multilass lassi ation problems. The issue of de omposing a Nlass lassi ation problem into a set of 2lass lassi ation questions is dis ussed. In parti ular, the te hnique for normalizing the outputs of several SVMs is presented. Based on these te hniques, support ve tor lassi ers for the re ognition of Western handwritten apitals are realized. Comparisons to several other lassi ation methods are also presented. 1 Introdu tion Support ve tor ma hines (SVMs) are primarily designed for 2lass lassiation problems1. Although SVMs a hieve substantial improvements over the urrently best performing methods and behave robustly over a variety of di erent learning tasks when a problem is treated as a binary lassi ation problem2;3, the appli ation of SVMs to multilass lassi ation problems is still a hallenge. Whereas in theory, the ombination of n SVMs an be used to solve a N lass (N > 2) lassi ation problem, su h a pro edure requires some are when applied to pra ti al problems4. In this paper, the issue of de omposing a N lass lassi ation problem into a set of 2lass lassi ation questions is dis ussed. For ombining the output of a set of SVMs, it is required that their outputs are normalized. In this paper, we will address the normalization of SVMs' output, testing the proposed te hnique on the lassi ation problem of simple bitmaps of Western handwritten apitals. Also, the use of several other lassi ation methods, su h as the Nearest Neighbor (1NN), k-Nearest Neighbor (kNN), Hidden Markov Model (HMM) and Multi-Layer Per eptron (MLP) is dis ussed in this paper. The experiments are performed with handwritten isolated upper ase English hara ters whi h are extra ted from the UNIPEN5 data base and onverted to pixel images. In this paper, se tion 2 is on erned with the basi idea of the Support Ve tor Ma hine (SVM) and the problems fa ed to the SVM when it is applied 167