RMO, Service de Recherche Technique de La Poste,

In this paper we introduce a new strategy for improving a discrete HMM-based handwriting recognition system, by integrating several information sources from specialized feature sets. For a given system, the basic idea is to keep the most discriminative features, and to replace the others with new ones obtained from new feature spaces. After evaluating the individual discriminative power of each single feature, the set is divided into two subsets: one containing the discriminative features, and the second the others. Considering feature classes in the nondiscriminative feature subset allows the specialization of new feature sets on specific problems. The application of this strategy to an existing system showed an improvement of 16 % in the recognition rate when a lexicon of 1000 city names was used.