Datenbestand vom 23. März 2024

Warenkorb Datenschutzhinweis Dissertationsdruck Dissertationsverlag Institutsreihen     Preisrechner

aktualisiert am 23. März 2024

ISBN 9783843951678

42,00 € inkl. MwSt, zzgl. Versand


978-3-8439-5167-8, Reihe Ingenieurwissenschaften

Simon Grosche
Non-Regular Image Sampling with Spatial Aperture

178 Seiten, Dissertation Universität Erlangen-Nürnberg (2022), Softcover, A5

Zusammenfassung / Abstract

Nowadays, almost all cameras are equipped with digital image sensors consisting of millions of pixels. Typically, the pixels have a square integration area and are regularly arranged next to each other covering the sensor surface. In the quest for higher spatial resolution, more pixels per area are used. This, however, reduces the number of photons reaching each pixel and increases the amount of data that needs to be acquired, processed, and stored. In presence of a fixed pixel budget, it has been shown that higher image quality can be achieved using non-regular sampling techniques such as quarter sampling and three-quarter sampling. This can be explained by the reduction of aliasing artifacts otherwise caused by conventional regular sampling. Quarter sampling and three-quarter sampling can be understood as special cases in the field of compressed sensing, where the task is similar to that of non-regular sampling.

In this thesis, the objective is to provide an image with the highest possible image quality for a limited number of measurements. To achieve this, the measurement process and the reconstruction process need to be considered. Regarding the measurement process, optimized point-like sampling strategies are presented for static and dynamic scenarios. For the latter, information about already measured data is taken into account during the selection of new sampling positions. Regarding the reconstruction, state-of-the-art sampling an reconstruction techniques are improved in terms of image quality and runtime. Finally, a new sensor layout consisting of T-tetromino pixels is proposed. It has the same number of pixels as a quarter sampling sensor and the same light-sensitive area as a (binned) low-resolution sensor. Together with data-driven reconstruction techniques, it outperforms classical single-image super-resolution by +1.62 dB.