A Comparison of Neural Network Algorithms and Preprocessing Methods for Star‐Galaxy Discrimination

We are interested in examining different artificial intelligence techniques for classifying astronomical objects. In this study we use two different neural networks that utilize supervised learning: learning vector quantization and back-propagation. The networks are trained to distinguish stars and galaxies using an example base of 17 × 17 pixel images consisting of 60 galaxies and 27 stars extracted from the first-generation Digitized Sky Survey. For each neural network we use four different preprocessing methods to create input vectors from the pixelized images. We also use as input the raw image data consisting of a 289 (17 × 17) point vector. Our results show that both networks are capable of distinguishing stars and galaxies, with back-propagation working somewhat better in most cases. We discuss the details of the preprocessing methods and which methods work better in which cases.