Deep Image Prior Neural Networks for interferometric image reconstruction

Abstract

We investigate Deep Image Prior neural networks for image reconstruction from sparse long-baseline interferometric data. The sky brightness distribution is parameterized as the output of an untrained U-Net-like convolutional generator, whose parameters are optimized independently for each data set by minimizing a loss defined on interferometric observables. The framework incorporates the forward model for visibility amplitudes, squared visibilities, and closure phases, while enforcing positivity and optionally including explicit regularization. The method is tested on Event Horizon Telescope observations of M87∗ and on simulated VLTI/GRAVITY observations of a time-variable asymmetric ring. For M87∗, the reconstruction recovers an asymmetric ring with a central brightness depression and a characteristic diameter of ∼ 45μas, consistent with horizon-scale synchrotron emission around the black-hole shadow. For dynamical VLTI imaging, coupling the DIP architecture to a continuous-time Dynamic Mode Decomposition model enables simultaneous recovery of multiple temporal frames from sparse (u, v) coverage, reproducing the counter-clockwise motion of the emission peak over a timescale of ∼ 27 days. These results indicate that untrained convolutional priors, combined with physically informed interferometric forward modeling and low-rank temporal dynamics, provide a flexible approach for static and time-resolved interferometric image reconstruction.