Date of Award

1995

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

Keywords

Computer Science.

Supervisor

Shridhar, M.,

Rights

info:eu-repo/semantics/openAccess

Abstract

The principal goal of this dissertation is to present several techniques for improving the performance of the handwritten numeral recognition system. There were two major tasks in this work. The first was to develop neural network approaches for improving the performance of classification. The second was to develop efficient fusion techniques for combining different classifiers. Apart from these major tasks, several feature extraction methods have been modified and some new features have been developed. By using a principal component analysis, a Bayes incremental learning neural network has been developed. This network is able to learn the data clusters as well as discriminants at a high speed. A multistage neural network architecture has been developed based on decomposition and localization strategy where different neural networks are used in different stages to perform different classification tasks. An evidence fusion technique based on the notion of fuzzy integral is utilized to combine classifiers. An algorithm for the dynamic assignment of source relevance is developed. By using the performance of each classifier on each class as well as the confusion information among classes, this algorithm effectively removes the discord error cases generated by individual classifiers. Experiments on handwritten numeral recognition are described. They show that multistage neural networks can generate a high recognition rate. These experiments also show that very low error rates with low rejection rates can be achieved by fusing several low performance classifiers.Dept. of Electrical and Computer Engineering. Paper copy at Leddy Library: Theses & Major Papers - Basement, West Bldg. / Call Number: Thesis1995 .C36. Source: Dissertation Abstracts International, Volume: 56-11, Section: B, page: 6220. Co-Advisers: M. Shridhar; M. Ahmadi. Thesis (Ph.D.)--University of Windsor (Canada), 1995.

Download

Share

COinS