Year

2021

Degree Name

Doctor of Philosophy

Department

School of Electrical, Computer and Telecommunications Engineering

Abstract

Over the past few years, deep learning (DL) has been shown to significantly im-prove the state-of-the-art in a wide variety of machine learning applications. It is a rich family of methods consisting of artificial neural networks (ANNs), optimiza-tion algorithms, and various unsupervised and supervised learning algorithms. With the remarkable advancement of GPU-based computational frameworks and the explosive availability of data, deep structured architectures with multiple-layer processing have recently made a strong renaissance. They can overcome the limitations of conventional shallow networks in working with the raw forms of natural multi-dimensional data.

Gabor filtering has attracted considerable research interest for neural net-works and visual recognition due to its biological evidence. It has been shown to analyze the oriented-frequency information occurring in an image region. Gabor filters are therefore particularly useful for texture representation and discrimina-tion. However, the main limitation of the existing Gabor-based methods is that they utilize only manually-designed Gabor parameters for feature extraction. Finding appropriate Gabor parameters for a certain problem is time-consuming and requires significant domain expertise. Furthermore, most Gabor networks are based on the traditional convolutional neural networks (CNNs) or the modulation technique with standard convolutional kernels. They are not full deep networks constructed from only Gabor filters with an end-to-end training algorithm.

FoR codes (2008)

0801 ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING, 0906 ELECTRICAL AND ELECTRONIC ENGINEERING

Download

Share

COinS
 

Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.