Recognition of Farsi handwriting strokes using profile HMM

This paper aims to stroke recognition, where the strokes are connected forms of cursive handwritten scripts, and in particular, we concern on recognition of Farsi handwriting strokes. In Farsi and some other writing systems, connected letters have special shapes that are often unrecognizable from their separated shapes. Despite that quite efficient algorithms have been developed for recognition of handwritten digits and disjoint letters, adapting these algorithms to stroke recognition is so arduous that development of a holistic approach is preferable. In this paper, we develop a method for Farsi handwriting recognition based on profile-HMM and study aspects of modeling the spatiotemporal features of handwriting strokes. The modular architecture of profile-HMMs provides a flexible framework for stroke modeling. Stroke shrinking and elongation are naturally modeled by the recurrent states and the silent states of profile-HMMs and make the model insensitive to writing speed and subtle slides. Our experimental results show that the profile-HMM is quite robust with respect to downsampling of the curve points, also is robust with respect to various settings in the training procedure. Our method correctly recognizes the main stroke of 90.8%, 98.5%, and 99.2% of handwriting samples, respectively in the top first, top five, and top ten hits.