A new deterministic annealing for image contextual classification

As an alternative to strict gradient descent-based procedures, we propose here a new deterministic annealing (DA) optimization approach for Bayesian-MRF contextual classification of images. The proposed DA builds on recent approaches that capture some of the power of the stochastic annealing optimization methods while reducing computational complexity via a deterministic approximation. It does not employ heuristic implementation to guarantee convergence. It uses the local Markov chains (LMC) modeling. Such local modeling is derived from the Markov model of simulated annealing. The update process of the LMC is obtained by using the modified Metropolis dynamics. Analytical guidelines for selecting the initial, critical, and final temperatures and the length of homogeneous LMC (number of iterations per temperature) for the annealing process are presented. The performance of the proposed approach is compared to other optimization methods. The method presents significant performance advantages compared to current methods for combinatorial optimization