Dynamic Handwriting Recognition Based on an Evolutionary Neural Classifier

摘要

At the present time, most of the classification problems have to deal with heterogeneous data presenting a strong variability, even within a class. It seems therefore relevant to substitute to the notion of class, the notion of sub-class, the latter regrouping a relatively homogeneous subset of examples. In order to generate these sub-classes (and models that are associated them) automatically, we developed an evolutionary neural classifier . At the beginning, it is made of as many networks as the number of classes of the problem. During the training, the number of networks evolves in order to modelize to the best the different sub-classes and to decrease the overall confusion rate between classes. An application of this classifier is the recognition of unconstrained dynamic handwriting: the multiplication of character models (called allographs) makes essential the automatic sub-class generation. Results, tested on some 25000 letters of the Unipen database are very encouraging.