A high-accuracy offline handwritten numeral recognition system.

Handwritten numeral recognition has been confronted with the problems of recognizing infinite varieties of patterns produced from writers with different writing habits, styles, and artistic flavors. As one of the most important topics in pattern recognition, there has been, and still is a significant performance gap between human beings and machines since the late 1960s. The primary objective of this research is to develop a high accuracy offline handwritten numeral recognition system. This thesis focuses on the architecture and performance improvement of handwritten numeral recognition systems through proper preprocessing, feature extraction, classifier design and combining different classifiers. Hybrid architectures of recognition systems are proved to be a very efficient method in recent research. This thesis proposes a multi-stage and multiexpert classification method integrated with complementary extracted features. It consists of a ruled-based classifier for one feature and neural network classifiers for all features. The final result is made from the fuzzy integral fusion of the outputs from the neural network classifiers. The experiments show that the approach achieves a better result.Dept. of Electrical and Computer Engineering. Paper copy at Leddy Library: Theses & Major Papers - Basement, West Bldg. / Call Number: Thesis2001 .Y44. Source: Masters Abstracts International, Volume: 40-06, page: 1596. Adviser: M. Ahmadi. Thesis (M.A.Sc.)--University of Windsor (Canada), 2001