The research for my bachelor thesis in august 2024 involved analyzing the performance of Convolutional Neural Networks (CNN) with custom filter initialization, especially for pulsar detection.
The goal is to improve pulsar detection by using customized filters for CNNs. With enhanced filter initialization, the CNNs are expected to converge faster and achieve greater accuracy.
The Deep Learning Framework PyTorch is used to create and train the networks. Therefore to be able to execute the programm, this framework is necessary (see https://pytorch.org/get-started/locally/ for further information).
I've got provided access to the Machine Learning Pipeline for Pulsar Analysis from PUNCH4NFDI, to generate synthetic data. The DM values are taken from https://www.atnf.csiro.au/people/pulsar/psrcat/ and a frequency range of 1.21 - 1.53GHz is used.
Example of Pulsar data:
The filters I've used (main usage in image processing):
- Prewitt
- Sobel
- Kirsch
- Canny-Algorithm
The CNN got trained on four different signal classes: Pulsar, No-Signal (no identified signal), BBRFI (Broadband Radio Interferenz) and NBRFI (Narrowband Radio Interferenz). By adding a layer at the beginning of the CNN for initial image processing, the CNN converged faster with custom filters.
To see the results for the different filters, please refer to results.
To read the thesis, please refer to Bachelorarbeit_Jad_Dayoub.pdf (in german).