Salient Object Detection With Higher Order Potentials And Learning Affinity

The human visual system can rapidly capture salient object in a cluttered visual scene by selective visual attention. This kind of capacity is also urgent to computer visual system for tackling the information overload problem. Therefore, numerous saliency models have been proposed to simulate human selective visual attention, and saliency detection have been widely applied to many computer vision tasks, e.g. image segmentation, object recognition, image retrieval and image compression.There are two contributions in this paper. First of all, we improve a novel graph-based salient object detection algorithm which exploits higher order potential to capture the cross-scale grouping cues instead of using a unified framework or naive multi-scale fusion. Higher order potentials enforce label consistency in image regions and in our framework we take the form of the Robust P" model. The application of higher order CRFs for saliency detection which integrate higher order potentials with conventional unary and pairwise potentials based on color and smoothness.What’s more, we investigate the importance of graph affinities in graph labeling. We take both local (spatial distribution) and nonlocal (feature distribution) priors into account and learn the pairwise similarity values in a semi-supervised manner, thereby obtaining a faithful graph affinity model. This nonlocal smooth prior and the well known local smooth prior can complement each other for salient object detection.Extensive experiments on three large benchmark datasets demonstrate the proposed method performs well when against the state-of-the-art methods in terms of accuracy.