Automatic Detection of Ship Based on Rotation Invariant RetinaNet

摘要

Ship detection in the aerial image is an active yet challenging in remote sensing image processing. Due to the difficulties of locating the arbitrarily rotated ships and the complexity of the background around the ship, the existing deep learning-based object detection algorithm is difficult to accurately identify the ship target. For multi-angle ships, the horizontal bounding box contains not only ship objects but also a lot of irrelevant backgrounds, they are learned by the deep con-volutional network as a learning object, and this leads to the common misalignment between the final classification confidence and localization accuracy. In this paper, we propose a new module called rotation RetinaNet (RRNet) to handle this problem. Firstly, we start from the popular one-stage RetinaNet approach, with ResNet50 as a basic network. Then, we apply a rotation head to RetinaNet to guarantee the rotation invariance of the model. Finally, we add an angular loss to the original loss so that the model can learn the angular offset of the bounding box. As a consequence, the proposed RRNet achieves high performance on the open-source DOTA datasets.