Restoration Networks for Large-Scale High Dynamic-Range Radio Interferometry Imaging

Plug-and-Play (PnP) algorithms solve inverse imaging problems by replacing handcrafted regularization with learned image priors encoded via deep neural networks (DNNs) within optimization algorithms. Existing PnP methods for radio interferometry (RI), such as AIRI, rely on Gaussian denoisers and achieve state-of-the-art reconstruction quality. However, the extreme dynamic range of the sought radio image and severe ill-posedness of the RI inverse problem require current PnP methods to be highly iterative, making them impractical for the large-data acquisition regimes of modern radio telescopes.

In this work, we propose a novel PnP algorithm for RI that extends the learned denoisers to more general restoration networks. This formulation enables the use of more powerful priors within a Half-Quadratic Splitting (HQS) algorithm. Combined with efficient Conjugate Gradient updates for the data-fidelity subproblems, these restoration priors substantially reduce the number of iterations required for convergence while delivering reconstruction quality competitive with AIRI and achieving approximately 4× faster reconstructions in simulations. We further validate the proposed approach on real observations, including small-scale narrow-field Very Large Array (VLA) data and large-scale wide-field MeerKAT observations.