Robust Background Exclusion for Salient Object Detection

Recently, investigating boundary prior to aid other low-level image cues, have gained great attention in salient object detection. Although the salient regions are mostly located in the image center, the inverse might not necessarily he true. In addition, such kind of center-bias assumption is very simple and fragile, especially when salient regions often touch the image boundary or the images are with complex background (e.g., small scale and high contrast patterns). To this end, this paper presents a new framework to address these two issues. First, we propose a robust background descriptor, called boundary connectivity, which is calculated to measure how heavily a region is connected to image boundary. This measure has an intuitive geometrical interpretation that are absent in previous saliency formulation. In order to reduce the effect of imprecise object boundary, we also propose a multiple layer over-segmentation framework to integrate multiple low-level cues, including our background measure, to highlight clean and uniform saliency maps. The experiment results demonstrate that our method achieves stateof-the-art results on MSRA-1000 datasets.