Salient object detection via spectral clustering

Detection of salient object is useful in many computer tasks. In this paper, we propose a novel spectral clustering (SC) based method to detect salient object. Image can be expressed as a graph which is composed by nodes and edges, where nodes are superpixels generated by segmenting algorithms and edge strengths are proportional to superpixels similarity. We jointly consider color distribution and spatial distance to measure the similarity between superpixels. Then SC algorithm is applied to cluster the nodes(superpixels) into two classes, one class for foreground and the other for background. It is reasonable that salient regions are far different from the background, so we can take salient object detection as a two-class clustering problem. Extensive experiments on a public database demonstrate that our model is not only easy to implement but also outperforms lots of recently proposed methods.