RCA: sparsity-based dimension reduction and super-resolution algorithm.

Python source code available here: rca.tar.gz

## The method

RCA (Resolved Components Analysis, **[1]**) aims at characterizing globally a space-variant PSFs field. Given a set of aliased and noisy images of unresolved objects (stars images from a telescope for instance), RCA estimates well-resolved and noise-free PSFs at the observations positions, in particular, exploiting the spatial correlation of the PSFs across the imaging system field of view (FOV). Let consider a set of p images of unresolved objects in a given instrument FOV and the corresponding ideal PSFs; these images are treated as column vectors and we note and . RCA minimizes , where accounts for the observed stars downsampling; this is done subject to several constraints over :

__Positivity constraint__: each PSF should be positive;__Low rank constraint__: each estimated PSF is forced to be a linear combination of a few number esimated of "eigen" PSFs; , with or equivalently, , and being low rank matrices;__Piece-wise smoothness constraint__: we can assume that the vectors are structured; promoting the sparsity of the "eigen" PSFs in an appropriate dictionary allows to capture the spatial correlations within the PSFs themselves;__Proximity constraints__: the more two PSFs are close in the FOV, the more they should be similar; the p values relative to the line of correspond to the contribution of the "eigen" PSF across the field of view; therefore the PSFs field regularity is enforced by constraining 's lines; specifically, each line is calculated as a sparse linear combination of spatial frequencies atoms: (see Fig.1).

Hence RCA solves the following problem:

The operator is set according to the resolution enhancement factor specified. Besides, this operator also accounts for the observed images subpixel offsets which are automatically estimated. is a user-chosen redundant dictionary. The frequency dictionary is automatically built based on the observations locations in the imaging instrument FOV. The sparsity parameters and as well are automatically selected. The number of "eigen" PSFs, r, has to be specified. It might be automatically reduced depending of the signal-to-noise ratio, the observations resolution and the sample size. A principal components analysis should provide the user a good first guess for r.

## Numerical examples

RCA was tested on set on 500 Euclid telescope simulated PSFs. The simulated observables were downsampled to Euclid telescope resolution and corrupted with a white gaussian noise. Four examples of recovered PSFs are given in Fig.2. In this example, 10 "eigen" PSFs were required. We compare the result to the PSFs obtained from the software PSFEx, using the same observations.

As it can be seen from these reconstructions, RCA handles the undersampling and the PSFs spatial variability, with both noise-robustness and a high visual quality.

## Article

**[1]**F. M. Ngolè Mboula, J.-L. Starck, K. Okomura, J. Amiaux, P. Hudelot.*Constraint matrix factorization for space variant PSFs field restoration*, Inverse Problems. Available here.