Harbor Detection in Large-Scale Remote Sensing Images Using Both Deep-Learned and Topological Structure Features

Abstract

Detecting harbors in large-scale remote sensing images has been a great challenge due to their relatively small areas and complex backgrounds. Furthermore, harbors usually contain different artificial or natural structures, which also increase the difficulty to detect them. A three-step harbor detection framework using both deep-learned and topological structure features is proposed in this paper. Firstly, coastal zones are found by Deep Residual Network (ResNet). Then, based on those coastal zones, different sizes of wharves are identified by Single Shot Multi-box Detector (SSD). Finally, semi-closed structure detector and bounding-box merging are implemented to obtain the final detection results. The experiments on large-scale remote sensing images have shown the proposed framework have better performances than those of both ResNet and SSD alone. The recall rate of the proposed framework is higher than that of ResNet about 15.9%, and that of SSD about 13.1%.