Smoothing of HMM parameters for efficient recognition of online handwriting

In this paper, we propose a novel approach to limited vocabulary recognition of unconstrained (mixed cursive) handwriting based on a hidden Markov model (HMM). Here, an input word sample is segmented into sub-strokes. However, instead of recognizing individual sub-strokes, we recognize the whole word sample as a basic unit. In this study, we consider both circular and linear features. The von Mises distribution is used to model circular feature components while the Gaussian distribution is used for the linear features. We implement fully connected non-homogeneous HMMs considering the enormous variability in the present handwriting style which often includes delayed strokes. Since such an HMM involves a large number of parameters, we implement a simple but novel approach to smoothing of HMM parameters to avoid possible over-fitting and poor generalization. Finally, in the proposed approach, we combine recognition results of two HMMs: one considering the input sample in the natural order and the other considering the sample in the reverse order. A major goal of the present study is to develop an efficient recognition scheme for online unconstrained handwritten words of Bangla, a popular Indic script.