Deterministic relaxation algorithms for edge detection and surface reconstruction

Summary form only given. The regularization framework has been extended to treat the joint optimization of an intensity process and an observable line process. The network of processes tries to satisfy soft and hard constraints as it relaxes to an equilibrium state. These constraints are usually (i) closeness to data, (ii) the smoothness criterion (the parallel to regularization), and (iii) constraints on line interactions. The functional to be minimized is nonconvex, and stochastic relaxation algorithms like simulated annealing can be used to obtain the global optimum. The approach is easily extensible to other early vision problems where discontinuities play an important role. The resulting task of optimization, however, is nonconvex. Several optimal techniques can provide good solutions. The problem can be stated in terms of finding the maximum a posteriori (MAP) estimate of a probability distribution or equivalently in terms of minimizing a potential function. Then the problem can be formulated in Bayesian terms (prior, degradation, and posterior distributions) and converted to one of minimizing a potential