Unsupervised saliency detection and a-contrario based segmentation for satellite images
Junbo Zhao, Shuoshuo Chen, Diyang Zhao, Hailun Zhu, Xiaoxiao Chen · 2013
In recent years, salient region detection techniques are widely used in image segmentation. The traditional image segmentation techniques primarily depend on human to label or mark the target areas interactively, which is far insufficient for real-time image processing. Therefore, in this paper we propose a new method of unsupervised saliency detection based segmentation, for high-resolution satellite images, which requires no manual interaction and prior knowledge of their content. Our proposed model of saliency at the considered pixel is a weighted average of dissimilarities between the pixel involved patch and the other patches. Moreover, we evaluated global and multi-scale contrast differences in order to extend the saliency calculation window to the entire image. To acquire an appropriate threshold for the remote sensing images segmentation, we apply a probabilistic a-contrario framework based on perception principle to measure the meaningfulness of such saliencies. According to the experimental results, our method is feasible and practicable for satellite image segmentation.