A Transductive Learning Method for Interactive Image Segmentation

The problem of interactive image segmentation is of great practical importance in image processing. Recently, a transductive framework has been developed which solves the problem from the perspective of manifold learning and shows promising results. In this paper, we extend this approach in two aspects. First, considering that the common interactive tools for user are broad brushes or region selection tools, it is hard to mark all the seeds accurately by hand. Our method is robust against noise, which releases the requirement of user and tolerant to a certain amount of user input faults. Secondly, we combine Our method with prior statistical information implicitly provided by user input seeds, result in higher accuracy. Experiments results demonstrate improvements in performance over the former methods.