Figure-ground segmentation using Tabu search

Many computer vision problems, such as figure-ground segmentation, simultaneous fitting of curves, selection of an optimal set of geometric primitives, can be formulated naturally as discrete optimization problems. The statement of these problems is relatively easy, but to find techniques that efficiently solve them constitutes a major challenge. In this paper the authors focus on figure-ground segmentation. The authors present a Tabu search strategy which is able to solve the discrete optimization problem associated with figure-ground segmentation in a very efficient way. The resulting deterministic algorithm outperforms the currently fastest known algorithm to solve this problem (mean field annealing) by two orders of magnitude in speed and in addition it consistently finds better optima.