Dissertations, Theses, and Capstone Projects

Date of Degree

9-2025

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy

Program

Computer Science

Advisor

Sos Agaian

Committee Members

Artyom Grigoryan

Shuqun Zhang

Louis Petingi

Subject Categories

Artificial Intelligence and Robotics | Computational Engineering | Signal Processing

Keywords

Quaternion neural networks; Retinex decomposition; Image dehazing; Image deraining; Adversarial robustness

Abstract

Computational Science and Engineering (CSE) combines math, statistics, computer science, and real-world knowledge to solve complex problems using computer models. This thesis presents Quaternion-Enhanced AI, which utilizes a special algebra called quaternions. Unlike real-valued numbers, quaternions are excellent at representing rotations and managing multichannel data effectively and efficiently, particularly in complex situations. We developed new quaternion-based neural networks to tackle challenging tasks in remote sensing, medical imaging, robotics, and computer vision when data quality is inadequate. Our work addresses problems where conventional methods often fail. The key contributions are a quaternion dehazing model that preserves colors accurately in foggy images; a rain removal network that requires 4× fewer parameters; a lightweight model that eliminates multiple weather effects like rain, snow, and fog; a unified quaternion system that addresses various weather challenges while enhancing object detection; a defense system against AI attacks that require minimal retraining; and a medical imaging model that improves contrast and assists doctors in identifying issues during capsule endoscopy. We tested our methods on real-world applications: clearing foggy images, removing rain and snow, safeguarding AI systems from attacks, and improving the clarity of medical capsule camera images. Extensive computer simulations demonstrate that quaternion networks are smaller, easier to comprehend, and more resilient to “bad” data than traditional methods. This work opens new possibilities for creating robust systems that understand geometry, leading to scalable and effective AI solutions for CSE applications, including remote sensing, medical imaging, self-driving systems, satellite imagery analysis, defense and surveillance, precision agriculture, underwater exploration, industrial quality control, and augmented reality.

Download

This work is embargoed and will be available for download on Wednesday, August 25, 2027

Share

COinS