Learning stable nonlinear cross-diffusion models for image restoration (Q2303272): Difference between revisions

Image restauration is successfully performed in the algorithm presented in this paper by the following three steps. The grey-scale image is formed as the two-dimensional solution of two coupled reaction diffuction equations with prescribed initial values. Now, in the first place, the equations are solved by finite difference methods. Then, secondly, the parameters of reaction and of the influence functions are computed by optimization using an augmented Lagrangian method and the Powell-Hestenes-Rockafellar function. The influence function is formed by a state-of-the-art radial basis function approximation with coefficients that are unknowns of the constrained optimization problem (as is the reaction parameter). Finally, the entries of the cross-diffusion matrix are computed using an artificial neural network. The stability of the numerical method for solving the differential equations is discussed in detail, and this is carried out both for the explicit and for the semi-implicit formulation. Many promising examples are given at the end of the article.