Salient Detection via Sparse Representation and Label Propagation

To improve the robustness of salient detection and increase the connection between the global information and local information, this paper proposes a salient detection method based on sparse representation and label propagation. First of all, to represent a data set succinctly and obtain a further relationship between the data, we define a new adjacency matrix that considers the regions located in the same subspace of data sets instead of the traditional definition of neighbor which share common boundary by using the sparse theory. Next, the weight matrix is computed by the similarity of the regions in the picture. And then, we select a part of boundary areas as background label. Finally, through weight matrix and background label, we adopt the label propagation to predict the label information of unlabeled region. We get the saliency map at last. Results on five benchmark data sets show that the proposed method achieves superior performance.