An improved classification for cursive script recognition

Abstract

We present a system able to recognize English words dynamically written on a digitizing tablet. It relies on a statistical approach, in which letters are first identified with a certain likelihood. A score is then given to each word in the dictionary for the final recognition. In order to optimize the allographs identification phase, a hierarchical classification algorithm allows to group the allographs corresponding to a given letter and a given size. Performance on a database of 5000 words and a dictionary of lo00 words are very good and the extension to multi-scriptors is possible thanks to a partial relearning.