New method of combining multiple classifiers for Chinese character recognition

Hong Guo,Xiaoqing Ding,Fanxia Guo,Youshou Wu

1997-01-01

Abstract:This paper analyzes the importance of combining multiple classifiers from the point of view of the information entropy theory. It also presents a new method, namely CRM (comprehensive recognition method), which can combine multiple classifiers. CRM is knowledge-based and takes account of the merits of different classifiers. As a result, CRM greatly improves the weakness of a single classifier for recognition of poor quality documents. The experiments show that compared with using single classifier, the error rate decreases by 74% and the rejection rate decreases by 27%. Additionally, an approach to select the rejection threshold of individual classifier based on a punishment function is put forward. The approach is flexible in determining the correct and error rates of each classifier and can avoid collecting a larger number of samples.