Bayesian Learning Based Visual Saliency Detection

This paper is to present a Bayesian learning based framework for visual saliency detection in natural scenes. Especially, for any point in the scene, this framework has considered whether it is salient or not; but previous methods by Bayesian learning seem not to do so. This framework includes two steps. First, the framework indicates that visual saliency is constituted with three main saliency modules. In a free-viewing manner, these main saliency modules are rarity, distinctiveness and central bias. Second, they are non-linearly combined for the final saliency map by a regularized neural network. The experimental results on two fixation datasets indicate that our framework outperforms other representative methods.