Hidden Markov models applied to on-line handwritten character recognition and signature verification

Abstract

Hidden Markov models are used to model the generation of handwritten, isolated characters. Models are trained on examples by optimization routines. Principally, optimization using the classical Lagrange multiplier method is possible, but in this work the so-called Baum-Welch optimization procedure is used. Then, given a model for each character in the vocabulary, unknown characters can be classified using maximum-likelihood classification. Experiments have been conducted, and an error rate of 6.6% over the English alphabet was achieved. Some experiments on the application of HMMs to signature verification are presented. Results indicate that HMMs provide a solid basis for a HMM-based signature verification system.