Hidden Markov models for couples of letters applied to handwriting recognition

This paper deals with handwritten word recognition using hidden Markov models (HMM) and presents a new solution to cope with problems of segmentation resulting from image preprocessing. This first step involves cutting an image of an isolated word into letters or pieces of letters called graphems. It builds a sequence of small images described by features which are the input of HMM. The image segmentation usually produces errors and lowers the results obtained by a recognition system based on a set of HMM models corresponding to the twenty-six letters of the alphabet. This paper proposes to extend the alphabet with models of couples of letters which are often badly segmented.