Cursive script recognition with time delay neural networks using learning hints

Abstract

In this paper, we present a system for cursive script recognition. It is built upon several procedures for preprocessing, a neural network to recognize individual characters in a word, and a hidden Markov model for the recognition of complete words. As neural network we use a time delay neural network that receives input from a sliding window, which scans an input word from left to right. The network is trained with the back-propagation algorithm which is expanded by position invariant learning. Because of the enormous need of computation resources we developed a method to include more information in the learning procedure. As a result only about a third of the cycles in the training are needed to achieve the same recognition results. Alternatively keeping the number of training cycles constant, an increase in word recognition rate of about 5% to a total of 87% can be achieved.