Date of Award

3-6-2009

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Operational Sciences

First Advisor

Kenneth W. Bauer, Jr., PhD.

Abstract

An automatic target classification system contains a classifier which reads a feature as an input and outputs a class label. Typically, the feature is a vector of real numbers. Other features can be non-numeric, such as a string of symbols or alphabets. One method of improving the performance of an automatic classification system is through combining two or more independent classifiers that are complementary in nature. Complementary classifiers are observed by finding an optimal method for partitioning the problem space. For example, the individual classifiers may operate to identify specific objects. Another method may be to use classifiers that operate on different features. We propose a design for a hybrid composite classification system, which exploits both real-numbered and non-numeric features with a template matching classification scheme. This composite classification system is made up of two independent classification systems.
These two independent classification systems, which receive input from two separate sensors are then combined over various fusion methods for the purpose of target identification.
By using these two separate classifiers, we explore conditions that allow the two techniques to be complementary in nature, thus improving the overall performance of the classification system. We examine various fusion techniques, in search of the technique that generates the best results. We investigate different parameter spaces and fusion rules on example problems to demonstrate our classification system. Our examples consider various application areas to help further demonstrate the utility of our classifier. Optimal classifier performance is obtained using a mathematical framework, which takes into account decision variables based on decision-maker preferences and/or engineering specifications, depending upon the classification problem at hand.

AFIT Designator

AFIT-DS-ENS-08-04

DTIC Accession Number

ADA495002

Download

Share

COinS