Date of Award

Spring 1995

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer and Information Science

First Advisor

Frank Y. Shih

Second Advisor

Peter A. Ng

Third Advisor

David Nassimi

Fourth Advisor

Daochuan Hung

Fifth Advisor

Edwin Hou

Abstract

Neural networks have been widely studied in a number of fields, such as neural architectures, neurobiology, statistics of neural network and pattern classification. In the field of pattern classification, neural network models are applied on numerous applications, for instance, character recognition, speech recognition, and object recognition. Among these, character recognition is commonly used to illustrate the feature and classification characteristics of neural networks.

In this dissertation, the theoretical foundations of artificial neural networks are first reviewed and existing neural models are studied. The Adaptive Resonance Theory (ART) model is improved to achieve more reasonable classification results. Experiments in applying the improved model to image enhancement and printed character recognition are discussed and analyzed. We also study the theoretical foundation of Neocognitron in terms of feature extraction, convergence in training, and shift invariance.

We investigate the use of multilayered perceptrons with recurrent connections as the general purpose modules for image operations in parallel architectures. The networks are trained to carry out classification rules in image transformation. The training patterns can be derived from user-defmed transformations or from loading the pair of a sample image and its target image when the prior knowledge of transformations is unknown. Applications of our model include image smoothing, enhancement, edge detection, noise removal, morphological operations, image filtering, etc. With a number of stages stacked up together we are able to apply a series of operations on the image. That is, by providing various sets of training patterns the system can adapt itself to the concatenated transformation. We also discuss and experiment in applying existing neural models, such as multilayered perceptron, to realize morphological operations and other commonly used imaging operations.

Some new neural architectures and training algorithms for the implementation of morphological operations are designed and analyzed. The algorithms are proven correct and efficient. The proposed morphological neural architectures are applied to construct the feature extraction module of a personal handwritten character recognition system. The system was trained and tested with scanned image of handwritten characters. The feasibility and efficiency are discussed along with the experimental results.

Share

COinS