Date of Graduation

12-2024

Document Type

Thesis

Degree Name

Master of Science in Computer Science (MS)

Degree Level

Graduate

Department

Electrical Engineering and Computer Science

Advisor/Mentor

Nakarmi, Ukash

Committee Member

Gauch, Susan E.

Second Committee Member

Zhang, Lu

Keywords

Computer Vision; Deep Learning; Image Reconstruction; Medical Imaging

Abstract

Inverse problems in computer vision involve reconstructing an original scene or image from incomplete, noisy, or indirect measurements. These problems are critical in tasks such as image denoising, deblurring, super-resolution, and lensless imaging, where the goal is to recover high-quality images from degraded or partial measurements. This thesis introduces novel approaches to address two specific real-world inverse problems: (1) image super-resolution in medical imaging and (2) lensless image reconstruction . In the first part of this work, we tackle the problem of image super-resolution in Magnetic Resonance Imaging (MRI). High-resolution MRI scans are often limited by hardware constraints, patient movement, and lengthy acquisition times. To address these challenges, we propose a fully unsupervised approach based on score-based diffusion models to super-resolve low-resolution MRI images. Unlike traditional methods that rely on simulated paired data, our model learns the underlying data distribution directly and reconstructs high-resolution MRI images from low-resolution inputs. Our method not only surpasses state-of-the-art supervised models in performance but also achieves faster sampling rates than current generative models for solving inverse problems. Additionally, we present an open dataset for training and benchmarking MRI super-resolution methods. In the second part, we propose an attention-based hybrid deep learning model for mask-based lensless imaging. Lensless imaging eliminates the need for expensive and bulky lenses, but traditional reconstruction algorithms suffer from slow convergence and poor perceptual image quality, while deep learning methods often introduce artifacts due to a lack of prior knowledge about the imaging model. Our approach integrates a conventional model-based optimization algorithm with an attention-based deep learning model. Experimental results demonstrate a significant improvement in perceptual quality compared to state-of-the-art methods.

Download

Share

COinS