Semester

Spring

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Nasser Nasrbadi

Committee Member

Matthew Valenti

Committee Member

Jeremy Dawson

Committee Member

Omid Dehzangi

Committee Member

Benjamin Riggan

Abstract

In recent years, generative adversarial networks (GANs) have shown great potential in advancing the state-of-the-art in many areas of computer vision, most notably in image synthesis and manipulation tasks. GAN is a generative model which simultaneously trains a generator and a discriminator in an adversarial manner to produce real-looking synthetic data by capturing the underlying data distribution. Due to its powerful ability to generate high-quality and visually pleasing
results, we apply it to super-resolution and image-to-image translation techniques to address vehicle detection in low-resolution aerial images and cross-spectral cross-resolution iris recognition. First, we develop a Multi-scale GAN (MsGAN) with multiple intermediate outputs, which progressively learns the details and features of the high-resolution aerial images at different scales. Then the upscaled super-resolved aerial images are fed to a You Only Look Once-version 3 (YOLO-v3) object detector and the detection loss is jointly optimized along with a super-resolution loss to emphasize target vehicles sensitive to the super-resolution process. There is another problem that remains unsolved when detection takes place at night or in a dark environment, which requires an IR detector. Training such a detector needs a lot of infrared (IR) images. To address these challenges, we develop a GAN-based joint cross-modal super-resolution framework where low-resolution (LR) IR images are translated and super-resolved to high-resolution (HR) visible (VIS) images before applying detection. This approach significantly improves the accuracy of aerial vehicle detection by leveraging the benefits of super-resolution techniques in a cross-modal domain. Second, to increase the performance and reliability of deep learning-based biometric identification systems, we focus on developing conditional GAN (cGAN) based cross-spectral cross-resolution iris recognition and offer two different frameworks. The first approach trains a cGAN to jointly translate and super-resolve LR near-infrared (NIR) iris images to HR VIS iris images to perform cross-spectral cross-resolution iris matching to the same resolution and within the same spectrum. In the second approach, we design a coupled GAN (cpGAN) architecture to project both VIS and NIR iris images into a low-dimensional embedding domain. The goal of this architecture is to ensure maximum pairwise similarity between the feature vectors from the two iris modalities of the same subject. We have also proposed a pose attention-guided coupled profile-to-frontal face recognition network to learn discriminative and pose-invariant features in an embedding subspace. To show that the feature vectors learned by this deep subspace can be used for other tasks beyond recognition, we implement a GAN architecture which is able to reconstruct a frontal face from its corresponding profile face. This capability can be used in various face analysis tasks, such as emotion detection and expression tracking, where having a frontal face image can improve accuracy and reliability. Overall, our research works have shown its efficacy by achieving new state-of-the-art results through extensive experiments on publicly available datasets reported in the literature.

Share

COinS