Discriminative Region based Two Stage System for Online Handwritten Character Recognition

This paper proposes a discriminative region based two-stage system for online handwritten Assamese character recognition. The first stage classifier employing hidden Markov model returns top-2 ranking classes. The second stage classifier based on support vector machine separates the top-2 ranking classes with better confidence. In literature, the second stage classifier is developed for each confusing character pair by considering the whole character pattern. Alternatively, in this work, the second stage classifier is designed considering the discriminative region (DR) of a confusing pair. It is to be noted that, a confusing pair share certain syntactic features and can introduce redundant information in the trained model. Therefore, by considering the features from DR, the separation can be enhanced. The proposed DR based two-stage system improves the recognition accuracy compared to the single stage and baseline two-stage system when evaluated on Assamese character database.