Can discrete Hidden Markov Models revolutionize online handwriting recognition? This paper introduces a fast HMM algorithm tailored for online handwritten character recognition. It enhances character recognition by leveraging a discrete HMM. Input strokes undergo discretization, leading to a simplified process for assigning initial state and state transition probabilities. The key to this approach lies in discretization, which simplifies state assignment and transition probabilities. The training phase avoids complete marginalization, and a normalized maximum likelihood ratio criterion guides the creation of new models to handle variations in stroke order and shape. Experiments on the Kuchibue database showcase the algorithm’s robustness to stroke number variations, with reasonable resilience against stroke order and shape differences. While memory usage is a concern for large character sets, density tying is suggested as a solution, paving the way for more efficient and accurate handwriting recognition systems. The work suggests that the algorithm is very robust against stroke number variations and has reasonable robustness against stroke order variations and large shape variations.
This article on handwriting recognition aligns with the International Journal of Pattern Recognition and Artificial Intelligence's focus on novel algorithms for intelligent systems. The proposed HMM algorithm contributes to the ongoing research in pattern recognition, specifically addressing challenges in online handwritten character recognition. The work's emphasis on robustness and efficiency is relevant to the journal's scope.