Estimation of super-resolved video dynamics
Résumé
In this work, we propose an efficient methodology for video super-resolution, that is, the recovery of a sequence of high-resolution images from its
low-resolution counterpart. The optimization problem associated to video super-resolution has several specificities which makes it particularly
challenging. A first barrier is the high-dimensionality of the problem, which derives from the extra temporal dimension and the unknown parametrization of the dynamical model characterizing the video. A second obstacle is the non-differentiability and the non-convexity of some of the terms of the cost function: the non-differentiability stems from the use of regularization terms of the state of the art (e.g., to enforce sparsity) whereas the non-convexity appears as soon as the motion describing the video is unknown.
In this paper, we propose an overall algorithmic framework to address the video super-resolution problem. Our approach is based on fast gradient
evaluation methods and modern optimization techniques for non-differentiable/non-convex problems. As a consequence, unlike previous work
in the field, we show that there exists a provably-convergent method estimating both the high-resolution image sequence and the underlying motion with a complexity linear in the problem dimensions. We assess the proposed optimization methods on videos of the MPI Sintel data set, known to be a challenging optical-flow benchmark.
Origine : Fichiers produits par l'(les) auteur(s)