Hidden Markov model with nonstationary states

In this paper we explore the nonstationarity of Markov model and we propose a nonstationary Hidden Markov Model (NSHMM) which is defined with a set of dynamic transitions probability parameters A(t)={a
ij
(t)} that depend on the time t already spent in state i. When compared to traditional models, this NSHMM is defined as generalization of the DHMM. The model was applied to on line recognition of handwritten arabie characters. The characters are represented by a radial sequence which is independent of translation, orientation and homothetie. The complete symbol-generation procedure includes sampling, size normalization and quantization phases. In the training, the model parameters are estimated with Baum-welch algorithm from a set of characters. Experiments have been conducted, and a good classification score has been obtained. The discrimination between characters models has been improved. Experiments showed that this approach can better capture the dynamic nature of arabie script.