Direkt zum InhaltDirekt zur SucheDirekt zur Navigation
▼ Zielgruppen ▼

Humboldt-Universität zu Berlin - Mathematisch-Naturwissen­schaft­liche Fakultät - Experimentelle Elementarteilchenphysik

Humboldt-Universität zu Berlin | Mathematisch-Naturwissen­schaft­liche Fakultät | Institut für Physik | Experimentelle Elementarteilchenphysik | H.E.S.S. | Theses | A Novel Approach to γ-Hadron Separation for H.E.S.S. Based on Convolutional Neural Networks

Tim L Holch (2016)

A Novel Approach to γ-Hadron Separation for H.E.S.S. Based on Convolutional Neural Networks

Master thesis, Humbold-Universität zu Berlin.

H.E.S.S. is an astroparticle experiment using imaging atmospheric Cherenkov telescopes to study cosmic γ-radiation in the GeV to TeV energy range. The vital task, as in most particle physics experiments, is an efficient discrimination of signal and background events in order to draw accurate conclusions from the performed observations. In the case of Cherenkov telescopes, the separation relies on significant differences in the recorded air-shower images. Computer vision techniques based on convolutional neural networks (CNNs) have shown to reach state-of-the-art performance in image classification tasks, promoting them to be explored as alternative γ-hadron separation technique for H.E.S.S.. The results of a pilot study on this topic are presented in this thesis, for which a deep CNN with four convolution layers and in total ∼ 1.6 · 10 6 tunable parameters was trained to separate images of Monte Carlo simulated γ- and hadron-showers. The developed CNN-classifier reaches a total classification accuracy of 97.41% on test images, where the application of the classifier on real H.E.S.S. data gives results in accordance with published H.E.S.S. findings. The general performance of the developed analysis shows significances and excess counts compatible with results obtained with standard H.E.S.S. analyses, thus suggesting CNNs to be applicable as γ-hadron separation method in H.E.S.S. data anal-ysis.