Arbitrary-Oriented Ship Detection via Feature Fusion and Visual Attention for High-Resolution Optical Remote Sensing Imagery

Ship detection for high-resolution optical remote sensing (HRORS) imagery is widely applicable to Navy construction, maritime transportation, maritime safety, port management, and many other fields. Optical remote-sensing images can easily be disturbed by thin cloud, light, and some complex environmental conditions, which makes the task of ship detection challenging. In this paper, we propose a new general detection network framework to recognize ships with arbitrary-orientation, in which we use a novel way to define the rotating rectangular box, remove the dependence on methods with angle parameters of the target. We propose a new feature fusion module (FFM) to integrate the semantic information, which enhances the feature extraction from the shallow network, and effectively improve the detection of small targets. We also employ one attention module (AM) to adaptively select meaningful features, which enhance the ability of feature extraction for complex environmental conditions. The multi-directional anchor size and length-width ratio are designed based on the clustering technique to improve the convergence speed and accuracy. Experiments on the public HRORS ship detection dataset HRSC2016 and the ship dataset SSD2020 collected from Google Earth have shown better performance compared with some conventional methods, which prove the feasibility and effectiveness of our proposed network architecture.