Algorithmes d'optimisation en grande dimension : applications à la résolution de problèmes inverses

摘要

An efficient approach for solving an inverse problem is to define the recovered signal/image as a minimizer of a penalized criterion which is often split in a sum of simpler functions composed with linear operators. In the situations of practical interest, these functions may be neither convex nor smooth. In addition, large scale optimization problems often have to be faced. This thesis is devoted to the design of new methods to solve such difficult minimization problems, while paying attention to computational issues and theoretical convergence properties. A first idea to build fast minimization algorithms is to make use of a preconditioning strategy by adapting, at each iteration, the underlying metric. We incorporate this technique in the forward-backward algorithm and provide an automatic method for choosing the preconditioning matrices, based on a majorization-minimization principle. The convergence proofs rely on the Kurdyka-L ojasiewicz inequality. A second strategy consists of splitting the involved data in different blocks of reduced dimension. This approach allows us to control the number of operations performed at each iteration of the algorithms, as well as the required memory. For this purpose, block alternating methods are developed in the context of both non-convex and convex optimization problems. In the non-convex case, a block alternating version of the preconditioned forward-backward algorithm is proposed, where the blocks are updated according to an acyclic deterministic rule. When additional convexity assumptions can be made, various alternating proximal primal-dual algorithms are obtained by using an arbitrary random sweeping rule. The theoretical analysis of these stochastic convex optimization algorithms is grounded on the theory of monotone operators. A key ingredient in the solution of high dimensional optimization problems lies in the possibility of performing some of the computation steps in a parallel manner. This parallelization is made possible in the proposed block alternating primal-dual methods where the primal variables, as well as the dual ones, can be updated in a quite flexible way. As an offspring of these results, new distributed algorithms are derived, where the computations are spread over a set of agents connected through a general hyper graph topology. Finally, our methodological contributions are validated on a number of applications in signal and image processing. First, we focus on optimization problems involving non-convex criteria, in particular image restoration when the original image is corrupted with a signal dependent Gaussian noise, spectral unmixing, phase reconstruction in tomography, and blind deconvolution in seismic sparse signal reconstruction. Then, we address convex minimization problems arising in the context of 3D mesh denoising and in query optimization for database management