Classification of radio galaxy images with semi-supervised learning

Abstract

The physical mechanism of galaxies lead to their complicated appearances, which could be categorized and require thorough study. Though millions of radio components have been detected by the telescopes, the number of radio galaxies, whose morphologies are well-labeled and categorized, is very few. In this work, we try to mind the features of radio galaxies and classify them with a semi-supervised learning strategy. An autoencoder based on the VGG-16 net is constructed first and pre-trained with unlabeled large-scale dataset to extract the general features of the radio galaxies, and then fine-tuned with labeled small-scale dataset to obtain a morphology classifier. Experiments are designed and demonstrated based on the observations from the Faint Images of the Radio Sky at Twenty-Centimeters Survey (FIRST), where we focus on the classification of three typical morphology types namely Fanaroff-Riley Type I/II (FRI/II), and the bent tailed (C-shape) galaxies. Compared to transfer learning on the same VGG-16 network, which was not trained with enough astronomical images and may suffer from a data-unseen problem, our semi-supervised approach achieves better performance at both high and balanced precision and recall.