Date of Graduation

5-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

John Gauch

Committee Member

Susan Gauch

Second Committee Member

Thi Hoang Ngan Le

Third Committee Member

Shengfan Zhang

Keywords

Attention Mechanisms, Generative Adverserial Networks, Image Classification, Machine Learning, Neural Networks, Semi-Supervised Learning

Abstract

Image classification is a sub-field of computer vision that focuses on identifying objects within digital images. In order to improve image classification we must address the following areas of improvement: 1) Single and Multi-View data quality using data pre-processing techniques. 2) Enhancing deep feature learning to extract alternative representation of the data. 3) Improving decision or prediction of labels. This dissertation presents a series of four published papers that explore different improvements of image classification. In our first paper, we explore the Siamese network architecture to create a Convolution Neural Network based similarity metric. We learn the priority features that differentiate two given input images. The metric proposed achieves state-of-the-art Fβ measure. In our second paper, we explore multi-view data classification. We investigate the application of Generative Adversarial Networks GANs on Multi-view data image classification and few-shot learning. Experimental results show that our method outperforms state-of-the-art research. In our third paper, we take on the challenge of improving ResNet backbone model. For this task, we focus on improving channel attention mechanisms. We utilize Discrete Wavelet Transform compression to address the channel representation problem. Experimental results on ImageNet shows that our method outperforms baseline SENet-34 and SOTA FcaNet-34 at no extra computational cost. In our fourth paper, we investigate further the potential of orthogonalization of filters for extraction of diverse information for channel attention. We prove that using only random constant orthogonal filters is sufficient enough to achieve good channel attention. We test our proposed method using ImageNet, Places365, and Birds datasets for image classification, MS-COCO for object detection, and instance segmentation tasks. Our method outperforms FcaNet, and WaveNet and achieves the state-of-the-art results.

Download

Share

COinS