Optimización binivel del parámetro de regularización con dependencia espacial del modelo de variación total generalizada para el filtrado de ruido en imágenes

Abstract

In this thesis, we study and solve a nonlinear bilevel optimization problem in function spaces. The goal is to determine the optimal spatially dependent regularization parameters in total variation (TV) and generalized total variation (TGV) image denoising models. Considering the spatial dependence of the parameters allows us to filter non-uniform noise in an image, which brings us closer to real situations where the type and distribution of noise are not known. We present some analytical results like the existence of solutions of the problem of parameter optimization, the Fréchet differentiability of the solution operator, which allows to prove the existence of Lagrange multipliers. The multipliers associated with the positivity constraints are regular Borel measured which are very difficult to compute. In order to overcome this issue, we proposed a Moreau-Yosida regularization, where the optimality system associated with the regularized problem was established and we prove that the solutions of regularized problems converge to the solution of the original one.