Deep Learning for Artifact Removal in Galaxy Images

In astronomical image acquisition, it is common to find artifacts and anomalies because the particularities of the studied objects (distance, light intensity, physical nature, etc.), as well as the acquisition process (instrumental aberrations, atmospheric turbulunce, etc.). Two of these aberrations are the Poisson noise and the effect of the point spread function (PSF).

Poisson noise occurs due to the oscillatory nature of light measurements by optical captation instruments. The low number of photons that the instruments capture means that this noise can be modeled using a Poisson distribution. It has the particularity of being closely correlated with the real image. On the other hand, the PSF models the response of an optical captation system to an input in the form of a Dirac delta, and it generates a blurring effect and a loss of spatial resolution. In the case of shift-invariant systems, the resulting image can be approximated as the convolution of the real image with the PSF.

Motivated for the recent advances in the field of Deep Learning for image reconstruction, we have built a solution based on convolutional neural network (CNN) for astronomic image aberrations removal.