Date of Award

12-1991

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Electrical and Computer Engineering

Abstract

Artificial neural network image segmentation techniques are examined. The biological inspired cortex transform is examined as a means to preprocess images for segmentation and classification. A generalized neural network formalism is presented as a means to produce common pattern recognition processing techniques in a single iterable element. Several feature reduction preprocessing techniques, based on feature saliency, Karhunen-Loeve transformation and identity networks are tested and compared. The generalized architecture is applied to a problem in image segmentation, a tracking of high- value fixed tactical targets. A generalized architecture for neural networks is developed based on the second order terms of the input vector. The relation between several common neural network paradigms is demonstrated using the generalized neural network. The architecture is demonstrated to allow implementation of many feedforward networks and several preprocessing techniques as well. Because of the limited resources and large feature vectors associated with classification problems, several methods are tested to limit the size of the input feature vector. A feature saliency metric, weight saliency, is developed to assign relative importance to the individual features. The saliency metric is shown to be significantly easier to compute than previous methods. Several neural network implementations of identity networks are tested as a means to reduce the size of the feature vectors presented to classification networks.

AFIT Designator

AFIT-DS-ENG-91-01

DTIC Accession Number

ADA243873

Comments

The author's Vita page is omitted.

Share

COinS