Abstract

We consider the problem of multi-label, supervised image segmentation when an initial labeling of some pixels is given. In this paper, we propose a new generative image segmentation algorithm for reliable multi-label segmentations in natural images. In contrast to most existing algorithms which focus on the inter-label discrimination, we ad- dress the problem of finding the generative model for each label. The primary advantage of our algorithm is that it produces very good segmen- tation results under two difficult problems: the weak boundary problem and the texture problem. Moreover, single-label image segmentation is possible. These are achieved by designing the generative model with the Random Walks with Restart (RWR). Experimental results with syn- thetic and natural images demonstrate the relevance and accuracy of our algorithm.