Astronomical image denoising by self-supervised deep learning and restoration processes

Tie Liu, Yuhui Quan, Yingna Su, Yang Guo, Shu Liu, Haisheng Ji, Qi Hao, Yulong Gao, Yuxia Liu, Yikang Wang, Wenqing Sun, Mingde Ding

Abstract: Image denoising based on deep learning has undergone significant advances in recent years. However, existing deep learning methods lack quantitative control of the deviation or error of denoised images. The neural network Self2Self was designed to denoise single images. It is trained on single images and then denoises them, although training is costly. In this work, we explore training Self2Self on an astronomical image and denoising other images of the same kind, a process that is also suitable for quickly denoising immense images in astronomy. To address the deviation issue, the abnormal pixels whose deviation exceeds a predefined threshold are restored to their initial values. The noise reduction is due to training, denoising and restoring and is, therefore, named the TDR method. With the TDR method, the noise level of solar magnetograms improved from about 8 to 2 G. Furthermore, the TDR method was applied to galaxy images from the Hubble Space Telescope, making weak galaxy structures much clearer. This capability of enhancing weak signals makes the TDR method applicable to various disciplines.