(An) HMM-based statistical framework for modeling on-line cursive script

Abstract

In this thesis, an HMM-based stochastic framework is proposed for the problem of on-line cursive script modeling. Basically, we use HMMs for modeling letters, the most natural unit chosen from the consideration of modeling difficulty and trainability. Then ligatures, the primary source of shape variability, are considered to include pen-up moves as well as pen-down drags. Clustered into a number of sets based on their position and left, right context, they are modeled as separate entities which in turn enable us to model both discrete and cursive unconstrained script. By concatenation of letter and ligature HMMs according to writing order and character (or word) composition rules, we are lead to a network model for on-line handwriting. When the design principle is extended to exploiting the sentential or multingual knowledge or more, a similar construction of HMM network is realized over a consistent framework. Given the network, the stochastic inference over the model is formulated into that of finding the most likely path. A recognizer is realized if the search is made for the best path given an observation sequence. A script generator is realized if the search is made for the best observation sequence from a given path. Finally, when the search is performed over a static handwriting image for the best sequence of observations and states, we can recover the pen trajectory that can aid or even serve off-line static image recognition. The search algorithms we have presented for those applications are based on the same technique of dynamic programming, and they are efficient and practical for real time applications. The capability of traditional HMM is often questioned because of its incorrect state duration modeling. As this can cause some problem when the the duration behavior is directly concerned, a new type of HMM called nonstationary hidden Markov model (NSHMM) is proposed in theoretical terms as a generalization - in descriptive power - of standard HMM; i...