Updating initial labels from spectral graph by manifold regularization for saliency detection

This paper presents a novel saliency detection method via updating initial labels from spectral graph in a semi-supervised learning (SSL) framework. For updating labels efficiently with graph-based SSL, two principles generally should be considered. The first one is that the updated labels should not change too much from their initial assignment. The second one is that the updated labels should not change too much between similar samples. To follow the first principle, the biggest eigenvector of Laplacian matrix, which contains rich contrast between background regions and salient regions, is employed to obtain the initial label vector. To follow the second principle, a new graph construction scheme, in which only boundary samples with similar features can be connected with each other, is proposed to reduce the geodesic distance in graph. Then a graph-based manifold regularization framework is exploited to update the label vector for separating salient samples from non-salient samples. A refinement function cooperating with an activation function is further presented for saliency optimization. Experimental results show that the proposed method achieves competitive performance against some recent state-of-the-art algorithms for saliency detection.