Abstract

The main purpose of this thesis is to discuss data-driven approaches to solve inverse problems in image reconstruction. In the Bayesian framework, the image prior serves as a regularizer in the computation of a maximum-a-posterior estimation of the reconstructed image. Classical image priors include Gaussian random space(e.g. Tikhonov regularization) or Besov prior (e.g. Total Variation regularization). Inspired by generative adversarial networks, a critic (discriminator) can serve as a regularizer, because of its capability of distinguishing the distribution of the ground-truth images from the distribution of the naively reconstructed images with classical regularization functional. Another data-driven approach, regularization by denoising (RED), provides a flexible and effective way to combine the state-of-the-art denoisers and model-based methods with a variety of optimization strategies to solve the inverse problem. Unlike traditionally hand-crafted regularizers, the data-driven regularization has the potential to learn an optimal regularizer from the data. In this thesis, we will consider two widely used linear forward models, and two data-driven approaches to solve inverse problem: adversarial regularizer and regularization by denoising.

Committee Chair

Umberto Villa, Ulugbek Kamilov

Committee Members

Umberto Villa Ulugbek Kamilov Netanel Raviv

Degree

Master of Science (MS)

Author's Department

Computer Science & Engineering

Author's School

McKelvey School of Engineering

Document Type

Thesis

Date of Award

Spring 5-21-2021

Language

English (en)

Download

Included in

Engineering Commons

Share

COinS