| 英文摘要 | As a branch of pattern recognition, the character recognition field has seen tremendous progresses during the past decades. The recognition of single characters has reached sufficiently high accuracies, and has been applied to various areas like character input for personal computers, mobile phones, PDAs, and so on. In many cases, however, people writes characters continuously, and requires new technologies for recognizing sentences or even whole paragraphs. Character string recognition, i.e., the recognition of a text line or a sentence, is the essential problem of handwritten text recognition. Although many research works of character string recognition have been done in the past few years, the current performance of character segmentation and recognition is still far from satisfaction for practical applications. The major difficulty of character string recognition is that the characters are hard to be segmented before they are recognized. A general solution is to integrate character segmentation and recognition in a combinatorial optimization framework. This dissertation proposes some strategies to improve the accuracy and search efficiency of character string recognition, and applies the methods to online handwritten Japanese character string recognition. The main contributions of this dissertation are as follows: First, we propose a machine learning based approach for modeling geometric context, to be used together with character recognition scores and linguistic context for evaluating search paths in integrated segmentation and recognition of handwritten character strings. By modeling both unitary geometry information (single-character geometry) and binary geometry information (character-pair geometry), the character segmentation and recognition accuracies were improved significantly. Second, this dissertation aims to improve the search efficiency of optimal segmentation-recognition path in character string recognition system. We implement and compare two search modes: character-synchronous search and frame-synchronous search, under normalized and un-normalized path scoring criteria. Also, we propose a modified path criterion which satisfies the principle of optimality such that the DP (dynamic programming) algorithm can be used to search for the optimal path. Third, due to the discrepancy of measurement ranges of different scores (character recognition, geometry, language) evaluating the candidate paths, it is difficult to tune the weighting parameters in the path criterion. We use confidence transformation methods to convert the (similarity/distance) scores into probabilities, such that the tuning of weighting parameters becomes easier. |
修改评论