Further Optimizations for the Chan-vese Active Contour Model

When a Chan-Vese active contour model is implemented in a framework that does not solve partial differential equations, we show how the mean pixel intensity inside and outside the curve can be updated efficiently. We reduce each iteration of the Chan-Vese active contour by O(n), when compared to an approach whereby the mean pixel intensities are recalculated for each iteration by looping over the entire image. After implementing the Chan-Vese active contour in the Shi-Karl level-set framework that does not solve partial differential equations, we show that the active contour may be trapped in an idempotent cycle, and we introduce a new stopping criterion to deal with this situation, thereby eliminating wasted computation cycles. Finally, we optimize the regularization cycle in the Shi-Karl framework, by detecting when an additional execution of the regularization cycle has no effect on the active contour, and breaking out of the loop.