GFB-Net: A Global Context-Guided Feature Balance Network for Arbitrary-Oriented SAR Ship Detection

Recently, deep learning techniques have been successfully applied to SAR ship detection. However, SAR ship detection is still a challenging task. First, ship targets in SAR images are characterized by arbitrary orientation and densely arranged. Traditional horizontal bounding box based detection is extremely unfriendly to such densely aligned scenes and often results in inaccurate localization. Secondly, the ship targets in SAR images are multi-scale, which makes detection difficult. This paper proposes an arbitrary-oriented SAR ship detection network called global context-guided feature balance network (GFB-Net) to solve the above problems. First, we introduce an efficient Oriented R-CNN detector. Unlike the horizontal bounding box detector, the rotated bounding box based detector can locate the ship target more finely and reduce the background interference. Additionally, we propose a global context-guided feature balanced pyramid to improve the detection performance of multi-scale ships by balancing different levels of feature and learning context information. Finally, we conducted experiments on SSDD dataset, and the results show that the proposed GFB-Net has a better performance compared to other detection methods.