Multi Dimensional Tracking, Classification, and Training Using Phase Only Filter Pattern Classifier

 
 
 
 

Multi Dimensional Tracking, Classification, and Training Using Phase Only Filter Pattern Classifier

Show full item record


Title: Multi Dimensional Tracking, Classification, and Training Using Phase Only Filter Pattern Classifier
Author: Gudmundsson, Karl Solvi
Description: Gudmundsson, Karl Solvi, Ph.D., Engineering Ph.D. Program, Department of Elecrical Engineering, Wright State University, 2004. Multi Dimensional Tracking, Classification, and Training Using Phase Only Filter Pattern Classifier. The only independent variable in electronic computing is time. Optical computing, on the other hand, has inherently two degrees of freedom, the two variables that define a point in a plane. Optical systems always process information in parallel. Such a simple optical element as a lens is capable of performing such a difficult task as a Fourier transform. Therefore, when it comes to pattern classification, optical computing is an attractive option. Dimension reduction and pattern classification from multi sensor data is a challenging exercise. The accepted methods of today's pattern classification of such data require the use of expensive algorithms. Algorithms such as: principal component analysis, linear discriminant analysis, and the k-nearest neighbor to name a few. These methods require significant computation time and are not easily made parallel. The Phase Only Filter (POF) has been shown to be a powerful tool for tracking objects in a two-dimensional plane. In this research, a special tracking technique is developed to overcome weaknesses of the POF under noisy circumstances. The POF is generally implemented in the 2-D plane. However, the POF has neither been trained as a pattern classifier for one dimensional data, nor as a multi-dimensional data classifier. Methods are developed in this work to apply the POF to multi-dimensional pattern classification. Possible optical architecture is also proposed that will allow a parallel realization of the system. Moreover, POF equivalent neural network techniques are devised and implemented for pattern classification. Two level neural network is developed for the case of multi-class classification, and a method of training is developed.
Permanent Link: http://rave.ohiolink.edu/etdc/view?acc_num=wright1166212400
http://hdl.handle.net/2374.OX/19742
Date: 2004

Files in this item

Files Size Format View

There are no files associated with this item.

This item appears in the following Collection(s)

Show full item record