L. Condat (GIPSA-lab) : A new primal-dual splitting algorithm for convex optimization; application as a heuristic for super-resolution
Date(s) : 30/01/2015 iCal
14h00 - 15h00
Summary: A new splitting algorithm is proposed to minimize the sum ofnconvex functions, potentially nonsmooth and composed with linearnoperators. This generic formulation encompasses numerous regularizedninverse problems in image processing. The algorithm, whose weaknconvergence is proved, calls the individual gradient or proximitynoperators of the functions, without any inner loop or linear system tonsolve. The classical Douglas-Rachford, forward-backward andnChambolle-Pock algorithms are recovered as particular cases. In thensecond part of the talk (joint work with A. Hirabayashi, Kyoto, Japan),nI address the problem of recovering a spike train from noisy lowpassnmeasurements, through a reformulation as a structured low rank matrixnapproximation problem. Used as a heuristic for this nonconvex problem,nthe proposed algorithm yields state-of-the-art results.
Catégories Pas de Catégories
