Optimal Seeds Extraction and Locally Smoothed Label Propagation for Salient Region Detection

To improve the robustness of salient region detection, this paper proposes a salient region detection model based on optimal seeds extraction and locally smoothed label propagation. By the model, the first step is to calculate a refined background map and extract optimal background seeds by random sampling in the map. Then an object prior map is generated by fusing two different prior maps. The optimal foreground seeds are extracted by thresholding. Finally, locally smoothed label propagation is proposed to predict labels of other regions by treating previously obtained seeds as initial labels and final saliency map is generated according to the predicted labels. Both quantitative and qualitative evaluations on three widely used datasets demonstrate the superiority of the proposed model to other several state-of-the-art models.