Visual Saliency Detection via Prior Regularized Manifold Ranking

Yun Xiao, Bo Jiang, Zhengzheng Tu, Jin Tang

Abstract: Bottom-up saliency detection has been widely used in many applications, such as image retrieval, object recognition, image compression and so on. Manifold ranking (MR) model can identify the most salient and important area from an image efficiently. One limitation of the MR model is that it fails to consider the prior information in its ranking process. To overcome this limitation, we propose a new manifold ranking model, called prior regularized manifold ranking (RegMR), which uses the prior calculating by boundary connectivity and employs the foreground possibility in the first stage and background possibility in the second stage. We compare our model with fifteen state-of-the-art methods. Experiments show that our model performs well than all other methods on four public databases on PR-curves, F-measure and so on.