Convex Formulation for Total-Variation Parameter Learning

Title

Convex Formulation for Total-Variation Parameter Learning

Speaker

Enis Chenchene

Abstract

We present a new approach for data-driven tuning of regularization parameters for total-variation denoising. The proposed approach hinges on a specific proxy for the underlying bilevel problem, which admits a convex monolevel reformulation that can be efficiently solved with a new conditional-gradient-type method. We show numerical experiments and open avenues for promising extensions.

Bio

Enis is currently a university assistant at the University of Graz. His primary research interests lie in splitting methods in convex optimization, optimal transport, and image processing. Before his current role, Enis earned his PhD in convex optimization at the University of Graz and worked on optimal transport in collaboration with MOKAPLAN in Paris.

When

Thursday March 3rd, 3pm

Where

Room 322 @ DIBRIS/DIMA, Via Dodecaneso 35, Genoa