文摘
The goal of saliency detection is to locate important regions in an image which attract viewers’ attention the most. In this paper, we propose a dynamic Bayesian model for saliency detection in which both Boolean-based and foreground-based models are exploited. First, a preliminary saliency map is constructed based on multi-channel Boolean maps, and a propagation mechanism is utilized to further modify the saliency map by learning a new weight matrix based on color and spatial structure information. Second, a foreground-based model based on foreground seeds from Boolean-based model is generated to detect salient pixels, and a better result is obtained by applying the edge map and a new weight matrix. Finally, pixel-level saliency is computed using a dynamic Bayesian framework. Both qualitative and quantitative evaluations on several benchmark datasets demonstrate robustness and effectiveness of our approach against state-of-the-art approaches.