It is generally agreed that the advantage of online character recognition methods with respect to off-line handwriting recognition mostly relies on the availability of dynamic information. The order of strokes and the other temporal information give good clues for solving off-line handwriting recognition problem. In this thesis, off-line handwriting recognition methods based on temporal information recovery is seriously discussed.
Among the previous work related to this issue, there is a good framework for recovering pen-trajectory based on stochastic information modeled in HMM. But the problem is that the pen-trajectory recovery algorithm does not always find the best solution. In this thesis, off-line handwriting recognition methods based on temporal information recovery is formulated to tree-searching problem, and several methods for improving the search algorithm are proposed to find the more feasible solution. The proposed methods are, increasing the width of search-bandwidth by preserving multiple candidates, utilizing the stochastic information more effectively by look-ahead, and integrating the structural and stochastic information. Additionally, a method for selecting start point of pen-movement is proposed.
Experiments shows that the proposed methods improves the recognition rates compare to the old one. Despite the improvements, the recognition rate is still somewhat poorer than the other off-line handwriting recognition algorithms.
