Dead-End Elimination as a Heuristic for Min-Cut Image Segmentation

We apply the dead-end elimination (DEE) strategy from protein design as a heuristic for the max-flow/min-cut formulation of the image segmentation problem. DEE combines aspects of constraint propagation and branch-and-bound to eliminate solutions incompatible with global optimization of the objective function. Though DEE can be used for segmentation into an arbitrary number of regions, in this paper we evaluate only the case of binary segmentation. We provide a runtime analysis and evaluation of DEE applied to two min-cut algorithms. Preliminary results show that DEE consistently reduces the search space for the Edmonds-Karp algorithm; tuning DEE as a heuristic for Boykov-Kolmogorov and other algorithms is future work.