Deep Learning VLBI Image Reconstruction with Closure Invariants

TL;DR

This paper introduces DIReCT, a deep learning method that reconstructs VLBI images directly from closure invariants, providing calibration-independent, high-fidelity results comparable to traditional algorithms without needing hyperparameter tuning.

Contribution

The work presents a novel deep learning approach for VLBI image reconstruction from closure invariants, bypassing the inverse transformation and calibration issues of prior methods.

Findings

Achieves median fidelity score > 0.9 on diverse images.

Insensitive to station-based corruptions and noise.

Comparable to state-of-the-art algorithms without hyperparameter tuning.

Abstract

Interferometric closure invariants, constructed from triangular loops of mixed Fourier components, capture calibration-independent information on source morphology. While a complete set of closure invariants is directly obtainable from measured visibilities, the inverse transformation from closure invariants to the source intensity distribution is not established. In this work, we demonstrate a deep learning approach, Deep learning Image Reconstruction with Closure Terms (DIReCT), to directly reconstruct the image from closure invariants. Trained on both well-defined mathematical shapes (two-dimensional gaussians, disks, ellipses, $m$-rings) and natural images (CIFAR-10), the results from our specially designed model are insensitive to station-based corruptions and thermal noise. The median fidelity score between the reconstruction and the blurred ground truth achieved is $\gtrsim 0.9$ even for untrained morphologies, where a unit score denotes perfect reconstruction. In our validation tests, DIReCT's results are comparable to other state-of-the-art deconvolution and regularised maximum-likelihood image reconstruction algorithms, with the advantage that DIReCT does not require hand-tuned hyperparameters for each individual prediction. This independent approach shows promising results and offers a calibration-independent constraint on source morphology, ultimately complementing and improving the reliability of sparse VLBI imaging results.