Building Extraction from Remotely Sensed Images by Integrating Saliency Cue

In this paper, we propose a novel two-step building extraction method from remote sensing images by integrating saliency cue. We first utilize classical features such as shadow, color, and shape to find out initial building candidates. A fully connected conditional random field model is introduced in this step to ensure that most of the buildings are incorporated. While it is hard to further remove the mislabled rooftops from the building candidates by only using classical features, we adopt saliency cue as a new feature to determine whether there is a rooftop in each segmentation patch obtained from previous step. The basic idea behind the use of saliency information is that rooftops are more likely to attract visual attention than surrounding objects. Based on a specifically designed saliency estimation algorithm for building object, we extract saliency cue in the local region of each building candidate, which is integrated into a probabilistic model to get the final building extraction result. We show that the saliency cue can provide an efficient probabilistic indication of the presence of rooftops, which helps to reduce false positives while without increasing false negatives at the same time. Experimental results on two benchmark datasets highlight the advantages of the integration of saliency cue and demonstrate that the proposed method outperforms the state-of-the-art methods.