Context dependent search in interconnected hidden Markov models for unconstrained handwriting recognition

Abstract

Unconstrained handwriting is especially difficult to recognize because of its variability and ambiguity. It is, however, not considered impossible if we can use various sources of knowledge. For example, a word is written letter by letter in a strict order. Viewing a handwritten word as an alternating sequence of characters and ligatures, we develop a circularly interconnected network of hidden Markov models that models the variabilities of handwritten English words of any length. In this network the task of recognition is defined by the process of finding the most probable path for a given input. The search is based on the Viterbi algorithm, which is a dynamic programming technique, that maximizes the alignment of a given input to all possible circular paths. The recognizer organized in this way carries out recognition and segmentation simultaneously. During the path search, we also consult a lexicon so that we can direct the search to a correct path. In the path search with the lexicon lookup, the interpretation of the initial part of a handwriting affects the interpretation of the following part. This implies that the search result depends on the search direction. It is very hard to decide the best search direction for a given input handwriting. This thesis analyzes the issue of directional sensitivity, and then proposes a dynamic backtracking technique called candidate refilling. Candidate refilling is a solution that remedies the problem of blocked path by backtracking dynamically to generate partial path candidates and replace the illegal candidates. Using this technique, the experimental recognizer is shown to achieve recognition rate over 99.42% of the correctness of the optimal recognizer efficiently. The advantage of candidate refilling method is more remarkable when the beginning parts of sample is written ambiguous. The complexity of search space is another difficulty. Since every point in a handwriting is potential segmentation point, we have to...