Unsupervised learning of character prototypes

Abstract

In the framework of handwritten word recognition, the use of characters extracted from words instead of written in isolation, is essential to train recognizers. We propose a segmentation method which relies on anchor points such as the ascenders or the descenders, but also on certain kinds of loops. We do not use a manually segmented prototype set to initialize our incremental learning process, but instead we use an a priori knowledge about the alphabet characters. This knowledge is introduced as the encoding of the descending movements of the pen and of loops. From a set of words written by the same writer, we evaluate the different possible segmentations for each word and use the ones superior to a certain threshold. In the beginning this threshold is rather high. At each step of the segmentation of the words, its value decreases in order to register new prototypes. The characters already accepted are used in the next steps. The confidence rate is maximum for 3 steps.