A hybrid radial basis function network/hidden Markov model handwritten word recognition system

A hybrid radial basis function network/hidden Markov model off-line handwritten word recognition system is presented. It is inspired from methods used originally in the field of automatic speech recognition. The hidden Markov model part of the system is in charge of modelling the alignment of letters onto segments produced by a rule-based explicit segmentation process. The role of the radial basis function networks is the estimation of emission probabilities associated to Markov states from the bitmaps of segments. It is shown that this system compares advantageously with a previous version using symbolic features as observations.