Combining Deformable Templates and Neural Networks for HCR

Abstract

This thesis describes a method of recognising handwritten digits by using a recurrent neural network to incrementally map a deformed character back to its undeformed template. In the deformable templates approach to handwritten character recognition, a character is described by a set of control points and spline segments are drawn through these points. A forward model of the distribution of characters is then obtained by adding a noise process at the location of these control points. This noise process should model the way characters are actually written. The inversion of the forward model yields a principled approach to HCR. However, this inversion is computationally expensive and even often intractable. For that reason the general neural network approach to HCR consists of directly mapping characters to classification. Nevertheless, this approach does not give any information about the character and such information would be relevant both to assess the reliability of the classification and to adapt quickly to the characteristics of a single writer. The aim of this project is, by relaxing control points to their home position, to compute the deformation of the character from the trajectory of the network. This method would resolve the problems described in the above paragraph and in addition would make it possible to estimate the few parameters of the forward model from a small training set if the network provides full inversion of it.