Several mathematical models and fast algorithms for image processing. (Quelques modèles mathématiques et algorithmes rapides pour le traitement d'images)

In this thesis, we focus on several mathematical models dedicated to low-level digital image processing tasks. Mathematics can be used to design innovative models and to provide some rigorous studies of properties of the produced images. However, those models sometimes involve some intensive algorithms with high computational complexity. We take a special care in developing fast algorithms from the considered mathematical models. First, we give a concise description of some fundamental results of convex analysis based on Legendre-Fenchel duality. Those mathematical tools are particularly efficient to perform the minimization of convex and nonsmooth energies, such as those involving the total variation functional which is used in many image processing applications. Then, we focus on a Fourier-based discretization scheme of the total variation, called Shannon total variation, which provides a subpixellic control of the image regularity. In particular, we show that, contrary to the classically used discretization schemes of the total variation based on finite differences, the use of the Shannon total variation yields images that can be easily interpolated. We also show that this model provides some improvements in terms of isotropy and grid invariance, and propose a new restoration model which transforms an image into a very similar one that can be easily interpolated. Next, we propose an adaptation of the TV-ICE (Total Variation Iterated Conditional Expectations) model, recently proposed by Louchet and Moisan in 2014, to address the restoration of images corrupted by a Poisson noise. We derive an explicit form of the recursion operator involved by this scheme, and show linear convergence of the algorithm, as well as the absence of staircasing effect for the produced images. We also show that this variant involves the numerical evaluation of a generalized incomplete gamma function which must be carefully handled due to the numerical errors inherent to the finite precision floating-point calculus. Then, we propose an fast algorithm dedicated to the evaluation of this generalized