Saliency detection in aerial imagery using multi-scale SLIC segmentation

Object detection in a huge volume of aerial imagery requires first detecting the salient regions. When an image is over-segmented by the superpixels, the latters will adhere to object boundaries, resulting in their shape deformation and size variation, which can be used as the saliency measure. The normalized Hausdorff distances from the inner pixels to boundary of the superpixels are then transformed to the posterior probability of saliency useful to build the saliency map and the salient region map. The method implemented by the multi-scale Simple Linear Iterative Clustering (SLIC) is simple without a priori knowledge on the object, computational efficient, robust to low image contrast, free of parameter tuning and is therefore suitable for aero-surveillance applications.