ausblenden:
Schlagwörter:
MAGIC
Zusammenfassung:
This thesis studies the possibility of using machine learning algorithms for the complete analysis of pixelized images produced by the MAGIC telescopes. The current approach reduces the information contained in the images down to a few parameters. Applying neural networks on the pixelized images can potentially use all the available information contained in the images and speed up the analysis. The focus of this thesis is the applicability of neural networks for the background rejection, direction reconstruction and image cleaning. Based on the PyTorch framework, neural networks that respect the hexagonal shape of the pixels were developed, trained and analysed with respect to the results of the current analysis method. The analysis shows that neural network based analysis techniques can be a valid alternative to the current approach and that can speed up the analysis and opens up the possibility of rejecting the background and analysing the images during operation on-site.
