Dissertation/Thesis Abstract

Image Super-Resolution using sharpened gradient profile prior
by Li, Ting, M.S., Southern Methodist University, 2012, 79; 1515308
Abstract (Summary)

The Super-Resolution (SR) image reconstruction has proved to be useful in many practical cases where multiple frames of the same scene can be obtained, including medical imaging, satellite imaging, and video applications. SR reconstruction could break the inherent spatial resolution limit of detectors in optical imaging systems by signal processing technics. Usually it is a numerically ill-posed problem and requires complex computation.

To solve this ill-posed problem, regularization is applied. In other words, prior information is combined in the reconstruction to make the problem well-posed. A variety of priors have been proposed, such as smooth prior, total variation prior, and edge statistics prior.

In this thesis, a novel image prior, which is called gradient profile prior, is adopted. This prior describes the shape and the sharpness of the image gradients. Based on the property of such gradient profile prior, the gradient transformation is implemented to obtain the sharpened image gradients. These image gradients can be used as a constraint in image super-resolution.

Such a prior was originally introduced for single-frame super-resolution; in this thesis, a new multi-frame super-resolution scheme is proposed to use such prior. This scheme is a two-step method. The first step is a basic L2 norm reconstruction without any prior, and then the gradient profile transformation is implemented on the first step result to get a sharpened gradient prior. Then the second step reconstruction is implemented, which uses an L2 norm for the fidelity part plus a sharpened gradient prior.

Identifying the proper gradient profile passing through one pixel is a crucial step in the gradient transformation. A new search method is proposed which takes into account both gradient magnitude and direction. Also the search is done within a neighborhood of pixels instead of tracing in a straight line. The modified method is faster and more accurate than the original method.

With a sharp and effective prior, we are able to produce better results. The two-step reconstruction keeps and enhances useful image features obtained by the regular reconstruction. Experimental results from both synthetic and real data have shown the effectiveness of this method.

Indexing (document details)
Advisor: Papamichalis, Panos E.
Commitee: Christensen, Marc P., Papamichalis, Panos E., Rajan, Dinesh
School: Southern Methodist University
Department: Electrical Engineering
School Location: United States -- Texas
Source: MAI 51/01M(E), Masters Abstracts International
Subjects: Electrical engineering
Publication Number: 1515308
ISBN: 9781267495884
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy