Saliency detection via local structure propagation

Abstract Saliency detection is a popular topic in computer vision, especially propagation-based method. This paper proposes a novel and effective coarse-to-fine saliency detection framework. In the coarse map stage, color spatial distribution map based on hue cue and central compactness rule is proposed, and integrated with texture boundary contrast information and background information to construct multi-prior coarse saliency map. In the refining stage, saliency values are updated under the comprehensive guidance of local structure propagation which is a novel algorithm to preserve local structural integrity during saliency propagation. With the global and local information, the detection procedure enhances the correctness of salient object gradually. Demonstrated in the extensive experiments on the public benchmark datasets, the performance of the proposed framework is superior to the state-of-the-art methods.