Abstract

Convolutional neural networks (CNNs) have been used for a wide range of applications in astronomy, including for the restoration of degraded images using a spatially invariant point spread function (PSF) across the field of view. Most existing development techniques use a single PSF in the deconvolution process, which is unrealistic when spatially variable PSFs are present in real observation conditions. Such conditions are simulated in this work to yield more realistic data samples. We propose a method that uses a simulated spatially variable PSF for the T80-South (T80-S) telescope, an 80-cm survey imager at Cerro Tololo (Chile). The synthetic data use real parameters from the detector noise and atmospheric seeing to recreate the T80-S observational conditions for the CNN training. The method is tested on real astronomical data from the T80-S telescope. We present the simulation and training methods, the results from real T80-S image CNN prediction, and a comparison with space observatory Gaia. A CNN can fix optical aberrations, which include image distortion, PSF size and profile, and the field position variation while preserving the source's flux. The proposed restoration approach can be applied to other optical systems and to post-process adaptive optics static residual aberrations in large-diameter telescopes.