Multi-initialisation segmentation with non-parametric minimum description length snake

Non-parametric polygonal snakes based on the minimum description length principle allow efficient segmentation of an object in a noisy image without a priori knowledge of the probability density functions of the image grey levels. However, at very low image contrasts, the segmentation quality can be degraded owing to convergence of the snake to local minima of the optimised criterion. This phenomena is analysed, a simple and robust optimisation strategy is proposed and a fast implementation is demonstrated.