人机结合的集成方法及其在字符识别中的应用

HUMAN-MACHINE COMBINED INTEGRATION METHOD AND ITS APPLICATIONS TO CHARACTER RECOGNITION

本文针对字符识别问题,提出一种人机结合的集成(integration)方法.这实质上是综合集成法(metasynthesis)的一种具体形式,即用基于监督学习的网络,把若干个识别系统所得到的结果,经过由教师指导的学习,综合各个系统的优点,进而获得比单个识别系统更好的结果.通过对四种自由手写体数字识别的分类算法加以综合集成,表明基于监督学习的网络集成法所取得的结果十分令人鼓舞.另外,当网络用于分类时,拒识阈值的确定往往要靠实验手段,为此本文还提出了一组用于确定阈值的公式,经实验证明,效果很好.

In this paper, a human-machine combined integration method is proposed aimed at solving the problems of character recognition. It is in essence a concrete form of metasynthesis, i.e., the results of several recognition systems are integrated by artificial neural network based on supervised learning. Thus, the advantages of individual systems are synthesized and a result which is better than any of them can be achieved. The result of the experiment is very exciting in which four free handwritten numeral recognition systems are synthesized by the proposed method. Moreover, when the network is used as classifiers, the rejection thresholds are usually obtained by the experimental method which is very tedious. To solve this problem, a group of formulas are presented. The effectiveness of these formulas is also demonstrated in the experiment.