AI for Large-Scale High-Dynamic Range Image Reconstruction in Radio Astronomy
DescriptionEndowing modern radio-interferometric (RI) telescopes with an acute vision requires solving large-scale high-dynamic range inverse problems for image formation. This calls for sophisticated computational imaging algorithms injecting an image model into the observed data. We discuss two AI-based solutions to break the barrier towards joint precision and scalability of RI imaging algorithms. The plug-and-play algorithm AIRI (arXiv:2210.16060) entails learning a prior image model by training a deep neural network (DNN) as a denoiser, and substituting it for the handcrafted, often sub-iterative, regularisation operator of an optimisation algorithm. AIRI is encapsulated in a parallel framework, equipped with automated functionalities for faceting its denoiser and decomposing the RI measurement operator into blocks. AIRI was demonstrated on gigabyte-sized data to deliver similar precision to its optimisation counterpart SARA, at reduced computational cost. Yet AIRI's highly iterative nature hinders scalability to extreme dimension. R2D2 (arXiv:2210.16060) is a learned version of the well-known matching pursuit algorithm. It delivers a reconstruction within few iterations, built as a series of residual images output by DNNs from data residuals. Preliminary results leveraging the same parallel measurement operator, suggest that R2D2 achieves similar precision to AIRI at a fraction of the cost, opening the door to real-time RI imaging.
TimeWednesday, June 2812:30 - 13:00 CEST
Computer Science, Machine Learning, and Applied Mathematics