Object or background: Whose call is it in complicated scene classification?

Scene semantic parsing is a challenging problem in the field of computer vision. Most approaches exploit low-level features to describe the whole scene. However, there is a large semantic gap between low-level features and high-level scene semantic. In this paper, a scene classification approach is proposed by exploiting semantic objects/materials of the background to reduce the semantic gap. The proposed approach can be divided three steps: First we construct two high-level semantic features (BCFs and BSLFs). Second, we design an approach to learn the prior probability of the Bayesian Networks from these two semantic features of training images. Finally, Bayesian Networks is used to achieve the goal of scene classification. Experimental results show that our approach achieves state-of-the-art performance on the task of scene classification compare with other approaches.