Exploiting global rarity, local contrast and central bias for salient region learning

• We build a saliency model combining global rarity, local contrast and central bias.
• We use the region information towards highlighting the whole object.
• We take a neural network based learning approach to combine the three factors.
• Our model outperforms 11 representative models to detect salient objects.
• Our model also well predicts eye fixations compared with eight classic models.

In this paper, we are to present a model that integrates and benefits from the global rarity, local contrast and central bias for saliency detection. Previous saliency works only consider one or two of them. Further, to avoid some inherent drawbacks of existing three factors, we first over-segment the image into many small coherent regions. And then, we exploit the self-information and regional de-noising, regional contrast and consistency, Gaussian function and regional averaging to get three new factors of global rarity, local contrast and central bias. Finally, we embed them into a nonlinear neural network to figure out their own contributions in saliency detection. Extensive experiments and comparisons illustrate the effectiveness of our saliency model with three new built factors.