Segmental K-means algorithm based hidden Markov model for shape recognition and its applications

In this paper, we propose a hidden Markov model (HMM) for object recognition on the basis of the shape of its boundary. The boundary can be looked upon as a sequence of certain elementary shapes. These elementary shapes are in fact, the states in the model. An HMM is quite appropriate in applications where the shape of the boundary varies significantly from one sample to another within a class. For the task of learning of the HMM parameters, the segmental K-means algorithm is used. This algorithm was basically developed for speech recognition and was not fully connected. Here it is modified as a fully connected HMM. The state distribution is assumed to be multivariate Gaussian. The proposed scheme has been tested on two databases of handwritten numeral and character shapes of Bangla script. The results are quite encouraging.