RANDOM WALKER WATERSHEDS: A NEW IMAGE SEGMENTATION APPROACH

We propose a new graph-based approach for performing a multilabel, interactive image segmentation using the principle of random walks. Using the random walk principle, given a set of user-defined (or prelabeled) pixels as labels, one can analytically calculate the probability of walking from each unlabeled pixel to each labeled pixel, thereby defining a vector of probabilities for each unlabeled pixel. By efficiently combining this vector of probabilities obtained for each unlabeled pixel, they can be assigned to one of the labels using the watershed algorithm to obtain an image segmentation. We present quantitative and qualitative results, comparing our new algorithm with the original random walker image segmentation algorithm.