New boosting algorithms for classification problems with large number of classes applied to a handwritten word recognition task

Abstract

Methods that create several classifiers out of one base classifier, so-called ensemble creation methods, have been proposed and successfully applied to many classification problems recently. One category of such methods is Boosting with AdaBoost being the best known procedure belonging to this category. Boosting algorithms were first developed for two-class problems, but then extended to deal with multiple classes. Yet these extensions are not always suitable for problems with a large number of classes. In this paper we introduce some novel boosting algorithms which are designed for such problems, and we test their performance in a handwritten word recognition task. © Springer-Verlag Berlin Heidelberg 2003.