Object detection based on BING in optical remote sensing images

High-resolution remote sensing images (RSIs) have been adopted in satellites, RSIs processing in satellites will enable new multimedia applications, such as situational awareness. In this paper, we have developed an object detection framework exploiting objectness measurement in which binarized normed gradient (BING) is used to detect some particular objects in RSIs more efficiently. Specifically, we (a) use BING to encode closed boundaries of objects and then obtain a group of candidate regions. The group can narrow down the search space of object detection, accelerating the traditional process of sliding window; and (b) employ a more robust feature descriptor. The feature is obtained by combining the pyramid of histograms of orientation gradients (PHOG) with Elliptic Fourier Transform so as to detect multiscale objects effectively; and (c) apply an SVM classifier to detect particular objects. Experiments show that the proposed framework has achieved an 85.1% average detection precision, 35.1% higher than the HOG-SVM using the same dataset.