New holistic handwritten word recognition and its application to French legal amount

This paper presents a holistic recognition of handwritten word based on prototype recognition. Its main objective is to arrive at a reduced number of candidates corresponding to a given prototype class and to determine from them the handwritten class to be recognized. The proposed work involves only an accurate extraction and representation of three zones namely; lower, upper and central zones from the off-line cursive word to obtain a descriptor which provides a coarse characterization of word shape. The recognition system is based primarily on the sequential combination of Hopfield model and MLP based classifier for prototype recognition yielding the handwritten recognition. The handwritten words representing the 27 amount classes are clustered in 16 prototypes or models. These prototypes are used as fundamental memories by the Hopfield network that is subsequently fed to MLP for classification. Experimental results carried out on real images of isolated wholly lower case legal amount bank checks written in mixed cursive and discrete style are presented showing an achievement of 86.5 and 80.75% rate for prototype and handwritten word recognition respectively. They confirm that the proposed approach shows promising performance results and can be successfully used in processing of poor quality bank checks.