Authors: F. Ngolè-Mboula
Language: C++
Download: sprite_v1.tgz
Description: SPRITE: sparsity-based super-resolution algorithm

SPRITE: sparsity-based super-resolution algorithm

The method

SPRITE (Sparse Recovery of InstrumenTal rEsponse, 1) aims at computing a well-resolved compact source image from several undersampled and noisy observations. SPRITE solves the succession of problems of the form

 \underset{\Delta}\min \frac{1}{2}\sum_{k=1}^n{\|\mathbf{y}_k-f_k\mathbf{D}\mathbf{H}_k(\Delta+\mathbf{x}^{(0)})\|_2^2}/{\sigma_k^2} +\kappa\|\mathbf{w}^{(l)}\odot\Lambda\odot\Phi\Delta\|_1 ; s.t. \Delta \ge -\mathbf{x}^{(0)},

with l=1..N .
The vectors \mathbf{y}_k are the low resolution (LR) observations. The scalars f_k account for possible luminosity differences between the LR images and \sigma_k^2 is the noise variance in the LR image k^{th}. The matrices \mathbf{H}_k account for the shifts between the observations and the matrix D is the downsampling operator. The vector \mathbf{x}^{(0)} is a rough estimate of the well-resolved image. The final image is given by \mathbf{x}=\mathbf{x}^{(0)}+\Delta_N, where \Delta_N is the minimizer of the N^{th} problem solved.

\Phi is a user-chosen redundant dictionary. The method relies on the prior that a suitable solution \Delta should have a sparse decomposition into the dictionary \Phi. Finally the inequality constraint (which is element-wise) insures that the final well-resolved image has positive pixels values.

It's important to note that the only parameters to be provided by the user are $Phi$ and $kappa$. The other parameters are automatically calculated.

The source code can be downloaded here: SPRITE package. The data and IDL codes used to perform the benchmark tests presented in 1 are available here: super-resolution benchmark.

Please cite the reference below if you use these codes in a publication. This is a preliminary version of the SPRITE package, which may be imperfect. Please feel free to contact us if you have any problem.


  • [1] F. M. Ngolè Mboula, J.-L. Starck, S. Ronayette, K. Okomura, J. Amiaux. Super-resolution method using sparse regularization for point spread function recovery, Astronomy and Astrophysics, 2014. Available here.