AstroSR: A Data Set of Galaxy Images for Astronomical Superresolution Research

Abstract

In the past decade, various sky surveys with a wide range of wavelengths have been conducted, resulting in an explosive growth of survey data. There may be overlapping regions between different surveys, but the data quality and brightness are different. The translation of data quality between different surveys provides benefits for studying the properties of galaxies in specific regions that high-quality surveys have not yet covered. In this paper, we create a data set for analyzing the quality transformation of different surveys, AstroSR, using the galaxy images from overlapping regions from the Subaru/Hyper Suprime-Cam (HSC) and the Sloan Digital Sky Survey (SDSS). In addition, we use superresolution (SR) techniques to improve the quality of low-resolution images in the AstroSR and explore whether the proposed data set is suitable for SR. We try four representative models: EDSR, RCAN, ENLCN, and SRGAN. Finally, we compare the evaluation metrics and visual quality of the above methods. SR models trained with AstroSR successfully generate HSC-like images from SDSS images, which enhance the fine structure present in the SDSS images while retaining important morphological information and increasing the brightness and signal-to-noise. Improving the resolution of astronomical images by SR can improve the size and quality of the sky surveys. The data set proposed in this paper provides strong data support for the study of galaxy SR and opens up new research possibilities in astronomy. The data set is available online at https://github.com/jiaweimmiao/AstroSR .

Keywords

• Galaxies
• Galaxy equators
• Convolutional neural networks
• Ground-based astronomy
• Optical astronomy
• Astronomy data modeling