Electronic Theses and Dissertations

Identifier

945

Date

2013

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Engineering

Concentration

Electrical Engineering

Committee Member

Robert Kozma

Committee Member

Mohammed Yeasin

Abstract

Artificial neural networks are inspired by the abilities of humans and animals to learn and adapt. Feed-forward networks are both fast and powerful, and are particularly useful for statistical pattern recognition. These networks are inspired by portions of the brain such as the visual cortex. However, feed-forward networks have been shown inadequate for complex applications such as long-term optimization, reinforced learning and image processing. Cellular Neural Networks (CNNs) are a type of recurrent network which have been used extensively for image processing. CNNs have shown limited success solving problems which involve topological relationships. Such problems include geometric transformations such as affine transformation and image registration. The Cellular Simultaneous Recurrent Network (CSRN) has been exploited to solve the 2D maze traversal problem, which is a long-term optimization problem with similar topological relations. From its inception, it has been speculated that the CSRN may have important implications in image processing. However, to date, very little work has been done to study CSRNs for image processing tasks. In this work, we investigate CSRNs for image processing. We propose a novel, generalized architecture for the CSRN suitable for generic image processing tasks. This architecture includes the use of sub-image processing which greatly improves the efficacy of CSRNs for image processing. We demonstrate the application of the CSRN with this generalized architecture across a variety of image processing problems including pixel level transformations, filtering, and geometric transformations. Results are evaluated and compared with standard MATLAB® functions. To better understand the inner workings of the CSRN we investigate the use of various CSRN cores including: 1) the original Generalized Multi-Layered Perceptron (GMLP) core used by Pang and Werbos to solve the 2D maze traversal problem, 2) the Elman Simultaneous Recurrent Network (ESRN), and 3) a novel ESRN core with multi-layered feedback. We compare the functionality of these cores in image processing applications. Further, we introduce the application of the unscented Kalman filter (UKF) for training of the CSRN. Results are compared with the standard Extended Kalman Filter (EKF) training method of CSRN. Finally, implications of current findings and proposed research directions are presented.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.

Share

COinS