COMPARISON OF NEURAL NETWORK TRAINING APPROACHES IN LIMITED DATA SCENARIOS

Earth-based observations of the night sky are significantly affected by noise generated by atmospheric motion. Various hardware and software solutions have been used to mitigate this problem. In our work, we focus on evaluating the performance of simple neural networks on data obtained from small telescopes. We test their noise reduction capabilities depending on network size and training method. We evaluate their performance using the precision and recall statistics of star detection in processed images. We show that for low-resolution data, using large networks (which were tested and proven to be efficient for data from large telescopes) can lead to worse results than using smaller networks.