AI detection of S/N<1 sources in infrared images: a deep learning algorithm developed for the AZT24 facility at Campo Imperatore Observatory

Imaging in the near-infrared is affected by a background signal coming from both the terrestrial atmosphere and the instrument itself, which plays an important role in limiting the instrument performances even when standard hardware solutions are applied - like the cryogenic cooling. Several extremely faint sources - which still produce relevant count levels - can therefore remain hidden under the noise, or else their weak characteristic peaks could be mistaken as residual noise peaks. In recent years, the development of increasingly sophisticated and performative deep learning techniques has been finding a number of applications in astronomical data handling and process. We present here a study aimed to identify below-the-noise (S/N<~1) sources in near-infrared astronomical images. We used a dataset of images in the J (1.25-micron), H (1.65-micron) and K (2.2-micron) bands, acquired with the SWIRCAM near-infrared camera mounted at the AZT24 telescope in Campo Imperatore observatory in the decade 1999 - 2008. Each image from a first subset has been compared with the corresponding, photometrically deeper image from the 2MASS catalogue, producing a set of positions of the sources in 2MASS. After built a Denoising CNN with a paired catalog of 2MASS clean images and artificially added-noisy images with a GAN, the SWIRCAM images have then been fed as input to the CNN, with the aim of identifying a pattern in the background around the missed astronomical sources. The CNN has proven to be effective in removing IR image noise in a more efficient way with respect to classical analytical denoising algorithms, leading to detect extremely low S/N sources, which have also been compared to the validated catalog. The algorithm can be potentially applied to images coming from any telescope, identifying all the sources below the noise and above the intrinsic detectability threshold of the detector. As such, it represents a powerful way to push the limiting magnitude of a telescope beyond the classical paradigm based on the signal-to-noise ratio only.