A bottom-up and top-down approach to using context in text recognition

Existing approaches to using contextual information in text recognition tend to fall into two categories: dictionary look-up methods and Markov methods. Markov methods use transition probabilities between letters and represent a bottom-up approach to using context which is characterized by being very efficient but exhibiting mediocre errorcorrecting capability. Dictionary look-up methods, on the other hand, constrain the choice of letter sequences to be legal words and represent a top-down approach characterized by impressive error-correcting capabilities at a stiff price in storage and computation. In this paper, a combined bottom-up top-down algorithm is proposed. Exhaustive experimentation shows that the algorithm achieves the error-correcting capability of the dictionary look-up methods at half the cost.