Deep Learning-based Image Deconvolution for the Next-Generation Astronomical Surveys

In the coming decade, large surveys, such as the Vera C. Rubin Observatory and Euclid, will cover vast areas of the sky by capturing millions of objects in multiple frequency bands. With such a wealth of galaxy images at our disposal, we stand to gain valuable insights into their origins and evolution. There is a great deal of phenomena in the fields of galaxy formation and evolution that require further understanding. These include inflow-outflow mechanisms, the position of galaxies in the cosmic web, their evolution, the importance of mergers, star formation, and the coevolution of galaxy morphology and star formation activity as a function of their place in the cosmic web. Additionally, there is a need to understand the morphological transformation as a function of environments, from voids to filaments to galaxy clusters, and answer questions such as when it begins. To achieve this, we must trace back the physical processes at play with look-back time. Finally, we also require high-resolution 2D mass maps.

This can only be made possible if we have access to clear, high-quality images. However, multiwavelength images over a wide field of view are only available from ground-based telescopes which introduce imperfections. Typically, the atmosphere and instrumental optics introduce a blurring effect in the images, modelled by a Point Spread Function (PSF), and the sensor variations introduce noise. Therefore, there is a crucial need to develop image deconvolution algorithms that can efficiently recover the morphology of galaxies along with their components—including star-forming regions, disks, bulges, etc.—while generalizing well to different galaxy images with varying noise properties.

To that end, we investigate the performance of a Deep Learning-based method (SUNet) and a classical deconvolution algorithm (Firedec). Our Deep Learning-based approach, which is orders of magnitude faster than its classical counterpart, involves a Tikhonov deconvolution with a closed-form solution, followed by post-processing with a Swin Transformer UNet (SUNet). We test the methods on real ground-based images obtained with the FORS2 camera at the Very Large Telescope (VLT) in Chile and demonstrate how SUNet shows good generalization to images with entirely different noise properties than the training dataset. Finally, we quantify the deconvolved objects in terms of their size and the number of clumps grouped with respect to their disc colour. This technique can be used to identify structures in the distant universe from ground-based images.