Deterministic networks for image estimation using a penalty function method

Summary form only given. A novel technique for image estimation which preserves discontinuities is presented. Gibbs distributions are used for image representations. These distributions also incorporate unobserved discontinuity variables or line processes. The degradation model is also Gibbs, which yields a posterior Gibbs distribution. The authors are interested in the maximum a posteriori (MAP) estimate. This reduces to finding the minimum of a Hamiltonian (energy function). The authors use a penalty function approach to solve the problem. This permits identifying the line processes as neurons with a graded response. The penalty function method also permits incorporating 'hard' and 'soft' constraints into the problem. These typically involve constraints on line endings, inhibition of adjacent parallel lines, preservation of line continuity of corners, etc. The authors propose two algorithms to solve this problem; the conjugate gradient (CG) and the iterated conditional mode (ICM) algorithms. Both algorithms are amenable to implementation on 'hybrid' networks.<>