Abstract
Spectral Segmentation is one of the methods used extensively to separate group with different characteristics for image segmentation in recent years. In image spectral segmentation, we should build up a similarity matrix. In this paper, we propose a method to build up the similarity matrix. We focus on the graph design based on salience, which fits the habit of human vision. Use the notion of salience to design a graph so that backgrounds and meaningful regions are more compact. In the graph, the weight of the Affinity Matrix between background and regions can be distributed reasonably. The region which we expect is more likely to segment out and then we learn to get a Full Pairwise Affinities Matrix. Finally, we run spectral segmentation with our Full Pairwise Affinities Matrix by using the graph to get the segmentation result. Our results exhibit the improvements for objects with similar colors to the background so that some segmentation algorithms are usually hard to find out the boundary of objects. Moreover, we improve the problem of unexpected boundary at smooth surfaces, which is caused by spectral segmentation.