|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 organized in form of a matrix addressed by integers. However, there are many image processing operations, such as rotation or zoom, that produce pixels at non-integer positions (see Fig. 1). These pixels have to be resampled onto the regular grid so the resulting image can be stored, displayed or further processed.
Figure 1: Example of applications that yield pixels at non-integer positions.
At LMS, we have developed the frequency selective resampling algorithm (MFSR) that produces high quality results and outperforms other resampling techniques such as nearest neighbour approach or linear interpolation (see Fig. 2). MFSR relies on a set of initially estimated pixels called key points. These key points are iteratively refined in order to obtain the desired resampled image.
Figure 2: Example of resampling for a 15% zoom-in operation.
In this thesis, the impact of the selected key points on the resulting reconstruction is to be studied. This work focuses on analysing how the quality of the key points affect the quality of the final reconstruction. The student shall implement some of the existing resampling techniques for key points estimation and build a testing framework. The performance shall be tested using Matlab. The thesis is to be written in English.