Arbitrary-Oriented Ship Detection based on Deep Learning

This paper explores a ship detector from an aerial view, where we focus on predicting ship orientation and improving recall rate for dense objects. First, since ships appear as small objects in remote sensing images, we enhance the backbone network in feature extraction, we reconstruct a classification network as a feature extractor, and then cascade a feature pyramid network for feature fusion. Second, to reduce the adverse effect of post-processing on dense predictions, we use rotated rectangular bounding boxes to represent ships, while adding angle predictions to the dense head. Finally, in order to increase the difference of features at different angles, we propose a target encoding. For the same kind of ships with different orientations, under this encoding rule, other predictions of the network, such as width and height, will have the same regression target. Considering both speed and accuracy, our method performs better in ship detection than other general detectors.