FSFADet: Arbitrary-Oriented Ship Detection for SAR Images Based on Feature Separation and Feature Alignment

Multi-oriented objects widely appear in scene texts and optical remote sensing images, and thus rotation detection has received considerable attention. However, we observe that few arbitrary-oriented ship detection methods for synthetic aperture radar (SAR) images have been proposed before. The main reasons are the essential difference between SAR images and optical remote sensing images and the lack of labeled data for training rotation detectors. In addition, there also exist these problems of arbitrary orientation, large aspect ratio, and dense arrangement in SAR ship detection task. To address these above issues, an arbitrary-oriented ship detection method named FSFADet based on feature separation and feature alignment is proposed. Considering the lack of labeled SAR images, we establish a new SAR ship rotation detection dataset named SSRDD dataset, which is an important task when using arbitrary-oriented ship detection approaches for SAR data. A feature separation module (FS-Module) is introduced to enhance the ship object feature and weaken the background noise. Meanwhile, a refined network (R-Network) and a feature alignment module (FA-Module) are introduced to boost the SAR ship detection performance. Finally, the IoU-smooth L1 loss is introduced to the loss function to address the boundary problem. The simulation experiments show that the proposed method is superior to other arbitrary-oriented object detection methods.