A Self-Attention Based SRGAN for Super-Resolution of Astronomical Image

High-resolution (HR) astronomical image play a vital role in the development of scientific research, cosmic exploration, astronomy, and physics. In this paper, we propose a self- attention based generative adversarial network of astronomical image super-resolution (SR) aiming at the problem of low- resolution (LR) of astronomical imaging systems. We adopt SRGAN as the benchmark model and add self-attention, which captures more global dependencies and deepens the network for enhanced high-frequency feature representation. To achieve fast and stable training, the BN layer is deleted from the proposed networks. The Charbonnier loss is introduced as the loss function to handle outliers and improve SR performance. Experimental results demonstrate that the proposed method is able to reduce artifacts and obtains better performance in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) on the astronomical image testset.