Saliency detection based on seed propagation in a multilayer graph

This paper presents a saliency detection algorithm based on the seed propagation in graph representation of image. First, the image is divided into over-segmented regions and a graph of the connected regions (nodes) with color similarity (edge weight) is constructed. Second, saliency seeds and background seeds are extracted based on contrast and boundary prior respectively. The contrast prior is used to select the regions to be used as saliency seeds, because salient regions usually have high contrast. Meanwhile, the regions for the background seeds are extracted from image boundaries, based on the observation that salient objects are seldom located on the image boundaries. Finally, using a semi-supervised learning framework, both saliency and background seeds are propagated to the entire nodes and the saliency of each node is calculated from the propagated results. In addition, we introduce a multilayer graph to further improve the performance, which makes the saliency values within an object more uniform. Comparison with the state-of-the-art methods shows that our algorithm yields comparable or better results on several datasets in terms of precision-recall curve, ROC curve, F-measure and area under ROC curve, and also show visually plausible results.