Curvature point based HMM state prediction for online handwritten assamese strokes recognition

Hidden Markov Models (HMM) are used in handwritten strokes recognition task. The two design parameters of HMM are the number of states and number of mixtures in each state. There are two approaches for finding the number of states, namely, equal number of states and variable number of states. Since the shape of strokes will be different, variable number of states approach should be beneficial. This work proposes a curvature point detection based method to predict variable number of states for modeling a handwritten stroke. The proposed method selects appropriate points from a trace so that the portion between two consecutive points is modeled as an HMM state. Accordingly, based upon handwritten stroke shape complexity, the number of appropriate points selected will change and hence the number of states assigned to the corresponding stroke. In the proposed method, the number of states is proportional to the shape complexity of the given stroke as opposed to fixed in case of brute-force. The HMM based stroke recognizer consisting of 181 distinct strokes, was trained on a set of 52,977 examples collected from approximately 100 native Assamese writers. The evaluation was done on 43,828 examples collected from same users in different sessions. The experimental results demonstrate the benefits of the proposed technique over the brute-force method, especially in case of complex shape strokes.