Date of Award

1993

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

S. Sitharama Iyengar

Abstract

This dissertation addresses a fundamental computational strategy in image processing hand written English characters using traditional parallel computers. Image acquisition and processing is becoming a thriving industry because of the frequent availability of fax machines, video digitizers, flat-bed scanners, hand scanners, color scanners, and other image input devices that are now accessible to everyone. Optical Character Recognition (OCR) research increased as the technology for a robust OCR system became realistic. There is no commercial effective recognition system that is able to translate raw digital images of hand written text into pure ASCII. The reason is that a digital image comprises of a vast number of pixels. The traditional approach of processing the huge collection of pixel information is quite slow and cumbersome. In this dissertation we developed an approach and theory for a fast robust OCR system for images of hand written characters using morphological attribute features that are expected by the alphabet character set. By extracting specific morphological attributes from the scanned image, the dynamic OCR system is able to generalize and approximate similar images. This generalization is achieved with the usage of fuzzy logic and neural network. Since the main requirement for a commercially effective OCR is a fast and a high recognition rate system, the approach taken in this research is to shift the recognition computation into the system's architecture and its learning phase. The recognition process constituted mainly simple integer computation, a preferred computation on digital computers. In essence, the system maintains the attribute envelope boundary upon which each English character could fall under. This boundary is based on extreme attributes extracted from images introduced to the system beforehand. The theory was implemented both on a SIMD-MC$\sp2$ and a SISD machine. The resultant system proved to be a fast robust dynamic system, given that a suitable learning had taken place. The principle contributions of this dissertation are: (1) Improving existing thinning algorithms for image preprocessing. (2) Development of an on-line cluster partitioning procedure for region oriented segmentation. (3) Expansion of a fuzzy knowledge base theory to maintain morphological attributes on digital computers. (4) Dynamic Fuzzy learning/recognition technique.

Pages

148

Share

COinS