Numerous image processing problems rely on the partionnning of an image in different classes. Such classes can be represented by labels belonging to a set of discrete possible values, that can possibly be non ordered. The minimization of non convex energies with respect to functions taking their values in such sets is nevertheless NP hard. For that reason, most popular methods are only able to compute local minima, using linearization, multi-resolution, graph representation... In this talk, recent works offering an interesting alternative will be presented. They are based on the convexification of non convex energies and are applied to a large range of classical problems: segmentation, optical flow computation, inpainting... With such reformulation of the problem, convex optimization tools can be used to obtain fast and accurate estimations of global minima of the original energies.