Interactive image segmentation based on hierarchical graph-cut optimization with generic shape prior

Abstract

A new algorithm for interactive image segmentation is proposed. Besides the traditional appearance and gradient information, a new Generic Shape Prior (GSP) knowledge which implies the location and the shape information of the object is combined into the framework. The GSP can be further categorized into the Regional and the Contour GSP to fit the interactive application, where a hierarchical graph-cut based optimization procedure is established, for its global optimization using the regional GSP to obtain good global segmentation results, and the local one using the Contour GSP to refine boundaries of global results. Moreover, the global optimization is based on superpixels which significantly reduce the computational complexity but preserve necessary image structures; the local one only considers a subset pixels around a contour segment, they both speed up the system. Results show our method performs better on both speed and accuracy. © 2009 Springer Berlin Heidelberg.