An improved superpixel-based saliency detection method

Abstract

In this paper, an improved saliency detection method based on superpixel is proposed. First, the original image is segmented into a number of superpixels by simple linear iterative clustering, each of which has the consistent color and texture characteristics. Second, two different methods, namely, the sparse representation-based method as well as a center-surrounding idea-based approach, are applied to these superpixels to compute the initial saliency map and a center-surrounding map, respectively. Then these two maps are integrated in an additive way to obtain a modified saliency map. Compared to the initial saliency map, the modified one is more precise. Third, for the segmented superpixels, a normalized cut-based clustering method is used to cluster them into several clustering areas, and then the salient values in the same clustering area are averaged. Consequently, we can get a much more uniform saliency map. Experimental results show that, compared with the classical algorithms, the proposed method achieves a better performance since it can highlight the salient objects evenly and restrain the background clutters effectively.