Star-galaxy separation by the residual neural network algorithm

In this paper, the residual neural network (RNN) algorithm was used to classify pseudo-color images of stars and galaxies from Sloan  digital  sky  survey (SDSS), with features obtained directly from images．Images of galaxies and stars with spectral information were used as training and test sets．After training, accuracy rate on the test set can reach 98.23%, and recall rate 98.80%, indicating that the RNN algorithm can accurately classify images of galaxies and stars．The probability of being a star or galaxy given by the classifier is verified, and the probability can be used to evaluate the reliability of classification．This classifier can be applied to future sky surveys to further test its performance．