Итерационные процедуры обработки изображений на основе ациклического динамического программирования

A lot of image analysis problems lend themselves to a unified mathematical formulation as optimization problems, namely (min, +)labelling problems. It is known that these problems are NP-hard in the general formulation, but when the adjacency graph is a tree, they can be effectively solved by the tree-serial dynamic programming procedure. In this paper, it is considered an optimization method on the basis of Gauss-Seidel principle, which uses iterative reevaluation of groups of variables with tree-like neighborhood relation. We change a way of such variable aggregation from iteration to iteration, as a mean for increasing the robustness of the algorithm with respect to local extremes.