Assessing the quality of restored images in optical long-baseline interferometry

TL;DR

This study evaluates various metrics for assessing the quality of images reconstructed from optical long-baseline interferometry data, identifying the l1 norm as the most robust metric for automatic quality assessment.

Contribution

The paper systematically compares image quality metrics in optical interferometry, proposing the l1 norm as the most effective for automated assessment of reconstructed images.

Findings

Convolution with an effective PSF is essential for proper quality assessment.

The effective resolution exceeds naive expectations due to prior information and object nature.

The l1 norm is the most robust metric, enabling automatic quality assessment.

Abstract

Assessing the quality of aperture synthesis maps is relevant for benchmarking image reconstruction algorithms, for the scientific exploitation of data from optical long-baseline interferometers, and for the design/upgrade of new/existing interferometric imaging facilities. Although metrics have been proposed in these contexts, no systematic study has been conducted on the selection of a robust metric for quality assessment. This article addresses the question: what is the best metric to assess the quality of a reconstructed image? It starts by considering several metrics, and selecting a few based on general properties. Then, a variety of image reconstruction cases is considered. The observational scenarios are phase closure and phase referencing at the Very Large Telescope Interferometer (VLTI), for a combination of two, three, four and six telescopes. End-to-end image reconstruction is accomplished with the MiRA software, and several merit functions are put to test. It is found that convolution by an effective point spread function is required for proper image quality assessment. The effective angular resolution of the images is superior to naive expectation based on the maximum frequency sampled by the array. This is due to the prior information used in the aperture synthesis algorithm and to the nature of the objects considered. The l1 norm is the most robust of all considered metrics, because being linear it is less sensitive to image smoothing by high regularisation levels. For the cases considered, this metric allows the implementation of automatic quality assessment of reconstructed images, with a performance similar to human selection.