On-line recognition of handwritten characters applying hidden Markov models with continuous mixture densities

This paper presents an investigation of on-line recognition of handwritten characters applying hidden Markov models (HMMs) with a finite-state Markov chain and a set of output distribution functions which are mixture of Gaussian density functions. The problem of handwritten character recognition is modelled in the framework of HMMs. Some attention has also been focused on problems related to model training for continuous HMMs because it is generally a difficult task. An iterative model training process consisting of pseudo random model initialisation, a k-mean clustering algorithm based initial model estimation and model reestimation stages is proposed. For each character, two HMMs are obtained based on a sequence of training data. Characters are represented using directional angle vectors and radius distance vectors. The recognition is performed using Viterbi algorithm. Experimental results based on handwritten digits and lowercase letters from three subjects are presented which show that HMM technique is very potential for handwriting recognition