Bottom-up model of visual saliency: A viewpoint based on efficient coding hypothesis

This paper proposes a novel bottom-up saliency model based on the mechanism of the early vision system. A relationship between the efficient coding theory and bottom-up saliency map in primate visual cortex is established. In this paper, we make a distinction of neural response between activated and inactivated by sparse coding, and define the saliency as uncertainity of internal representation. Beyond the definition of saliency, our model also accounts for the issue of why we need such a saliency map. Finally, we test this model on artificial images such as psychological patterns and two different scale datasets. Although it is only a simple model of bottom-up saliency, the experiment results show it outperforms other state-of-the-art methods.