Salient Region Detection via Low-level Features and High-level Priors

Humans have the capability to quickly prioritize external visual stimuli and localize their most interest in a scene. However, computational modeling of this basic intelligent behavior still remains a challenge. In this paper, we formulate salient region detection as a binary labeling problem that separates salient region from the background. A Conditional Random Field is learned to effectively combine low-level features with high-level priors. We use a set of low-level features including local features and global features. We use the low level visual cues based on the convex hull to compute the high-level priors. Experimental results on the large benchmark database demonstrate the proposed method performs well when against six state-of-the-art methods in terms of precision and recall.