Paper 13101-168
Machine learning and ai software for image improving and restoring
19 June 2024 • 17:30 - 19:00 Japan Standard Time | Room G5, North - 1F
Abstract
Disturbance removal in astronomical images is of paramount importance both for automated processing methods
and for “manual” inspection of images by humans. Disturbances in images comes in various forms and can be
introduced in the raw signal itself as either noise, when random processes are involved, or artifacts, due to signal
acquisition modality. Such is the case of signals acquired by earth rotation aperture synthesis where astronomical
observations are made by integrating multiple signals coming from arrays of geographically distributed radio
antennas. In this work we provide a dataset and present a deep learning nertwork for removing interferometry artifacts in
combined images generated by aperture synthesis. The model combines the capabilities of both U-Nets [1] and
Vision Transformers [2]. The CANDELS dataset was used as the seed dataset. The 16200×16200 images were
randomly sampled by extracting random crops of size 1024×1024. The model was trained on a split comprising 60% of the dataset (60,000 images), 20% (20,000 images) was used for validation and 20% (20,000 images) in testing achieving a PSNR of 18db and a SSIM score of 0.51.