ISBN 978-3-8439-2981-3

978-3-8439-2981-3, Reihe Ingenieurwissenschaften

Thomas Richter
Super-Resolution Techniques for Mixed-Resolution Multi-View Images

212 Seiten, Dissertation Universität Erlangen-Nürnberg (2016), Softcover, A5

Zusammenfassung / Abstract

Super-resolution approaches aim at increasing the spatial resolution of a given low-resolution input image. Basically, respective methods can be grouped into single-image and multi-image approaches, which can be further subdivided, depending on whether temporally or spatially neighboring images are used for super-resolution.

Within the scope of this thesis, super-resolution is conducted for the special case of mixed-resolution multi-view images, where a scene is captured from different perspectives with cameras providing various spatial resolutions. Using the concept of high-frequency synthesis, neighboring high-resolution views can be efficiently exploited. Under ideal conditions, remarkable gains can be achieved, by projecting the missing high-frequency part from a reference perspective onto the image plane of the target view. However, these ideal conditions are typically not fullfilled for practical scenarios, negatively influencing the achievable super-resolution quality.

This thesis deals with different challenges, arising in the context of super-resolution based on high-frequency synthesis. First, an estimation method is discussed, which aims at extracting a proper high-frequency part that, after projection, well fits the desired target view. Since the required high-frequency projection is based on depth information, the influence of inaccurate depth acquisition is conducted. As a consequence, a robustly extended high-frequency synthesis method is proposed. The simulation results show that, using this extension, high-quality super-resolution results can be obtained, even in the case of strong depth distortions or inaccurate multi-sensor calibration. In addition, dealing with remaining occluded areas is intensively investigated. Depending on the camera setup and the scene geometry, some image parts might not be visible in any of the reference views, leading to holes in the synthesized high-frequency image. To properly handle these unknown high-frequency parts, different strategies are proposed, clearly outperforming state-of-the-art reconstruction methods. However, even in the case of mixed-resolution arrays, high-resolution reference cameras are required. In order to avoid the need of high-resolution sensors, the concept of non-regular sampling is introduced. A reconstruction scheme for irregularly sampled multi-view images is discussed at the end of this thesis.