Restoration Networks as Priors for Large-Scale Radio-Interferometric Imaging

Modern radio telescopes such as MeerKAT, and soon the SKA, generate enormous amounts of data. Turning that data into high-fidelity images is a severely ill-posed inverse problem: the sky is sampled incompletely in the Fourier domain, and faint emission must be recovered next to sources up to a million times brighter. Learned image priors have brought radio-interferometric imaging to state-of-the-art quality, but at a price. Plug-and-Play (PnP) methods, which alternate a physics-based data-fidelity step with a learned denoiser, need thousands of iterations to converge. Faster end-to-end networks, on the other hand, are tied to the telescope configuration and image size they were trained for.
In this talk, I'll present a PnP-inspired algorithm that bridges the two. Building on recent advances in computer vision, it embeds image-restoration networks, trained to remove both noise and blur, as priors in a half-quadratic splitting scheme. It reaches high-quality reconstructions in far fewer iterations, without retraining for each telescope configuration. The result is reconstruction quality competitive with the state of the art at a fraction of the computational cost, demonstrated on real data from the VLA (Cygnus A) and on wide-field MeerKAT observations of the radio galaxy ESO 137-006.