CONTEXT-DEPENDENT SEARCH IN INTERCONNECTED HIDDEN MARKOV MODEL FOR UNCONSTRAINED HANDWRITING RECOGNITION

Abstract

Viewing a handwritten word as an alternating sequence of characters and ligatures, we proposed a circularly interconnected network of hidden Markov models to model handwritten English words of indefinite length. The recognition problem is then regarded as finding the most probable path in the network for a given input. For the search, Viterbi algorithm is applied with lexicon lookup. To overcome directional sensitivity of the path search, a back-tracking technique is employed that keeps plausible path candidates dynamically within limited storage space.