|Betreuer:||PhD Jan Koloda (Raum 01.178)|
|Hochschullehrer:||Prof. Dr.-Ing. André Kaup|
Digital images are commonly represented as regular 2D arrays, so pixels are organised in form of a matrix addressed by integers (i.e. rows and columns). However, there are many image processing operations, such as rotation or warping, that produce pixels at non-integer positions. Since it is not possible to directly display these pixels, they must be regularized and pixels on the regular grid have to be estimated. This regularization can be alternatively seen as a reconstruction problem where the unknown pixels on the regular grid are recovered using the available pixels at non-integer positions.
However, many of the image recontruction techniques assume that the input image is already a regular array and therefore cannot be applied. Some of the algorithms applicable to this non-integer scenario are based on triangulation. This typically yields noisy reconstructions with notable artefacts. In order to improve the reconstruction quality, denoising techniques can be applied afterwards.
In this thesis, the possibility to utilise visual features and available signal statistics for controlling the denoising strength is to be studied. Once a controlling mechanism is obtained, it shall be plugged into a Matlab framework in order to test the performance and compare it to already existing techniques. The thesis is to be written in English.