Neuro-Markovian hybrid system for handwritten arabic word recognition

Automatic reading of handwritten words is a difficult problem, not only because of the great amount on variations involved in the shape of characters, but also because of the ambiguities. Several factors permit to judge of the problem complexity. In this paper, a hybrid recognition system using neural networks and hidden Markov models is presented for reading bank checks. The word images are transformed into portions of characters called graphemes which are analysed based on their shape (geometrical and metrical features). The word is then coded by a sequence of observations similar to human perception. The results of the above step will be used at the recognition level, which is based on probabilistic models. Experimental results obtained on a database of 5000 samples are reported and compared.