Detection de contours et restauration d'image par des algorithmes deterministes de relaxation. Mise en oeuvre sur la machine a connexions CM2

Recently, a lot of algorithms minimizing a non-convex energy function have been proposed to salve low level vision problems . Different kinds of relaxation methods are available . The stochastic techniques, such as simulated annealing, asymptotically converge to the global minimum but require a high computational cost . Deterministic relaxation methods which are sub-optimal, give good results and are faster than the stochastic ones . In this palier, we focus on the parallel implementation of two deterministic algorithms for edge detection and image smoothing : the graduated nonconvexity (GNC) originally proposed by Blake & Zisserman and the mean field annealing (MFA) introduced by Geiger & Girosi and extended to anisotropie compound Gauss-Markov random fields by Zerubia & Chellappa . Both methods are based on a weak-membrane model and both algorithms are inherently serial : each step produces a pixel map which is taken as an input for the next step . For the GNC, we implement a checkerboard version of the successive over-relaxation (SOR) method to minimize the energy . For the MFA, we use an optimal step conjugale gradient descent .