Learning to detect salient region of image under weak supervision

Salient region of an image usually contains the crucial information for image analysis and understanding. Most conventional approaches learn the saliency by utilizing the low-level features, which ignore the participation of human. In this paper, we propose an effective and robust approach to detect the salient region of an image by combining the bottom-up and top-down cues. The proposed method not only consider the low-level attention features, but also take human into the loop for better understanding of human attention. Furthermore, we build an asymmetrical graph model to integrate these bottom-up and top-down cues into an energy function of saliency. A compact but exact saliency region can be obtained by minimizing posterior energy function. The compact constraint and global minimization manner of the asymmetrical graph cuts guarantee the good performance of saliency extraction. Extensive experiments demonstrate the proposed method is promising.