Scattering-Point-Guided RPN for Oriented Ship Detection in SAR Images

Ship detection in synthetic aperture radar (SAR) images has attracted widespread attention due to its significance and challenges. In recent years, numerous detectors based on deep learning have achieved good performance the field of SAR ship detection. However, targets same type always various representations under different imaging conditions, while types ships may a high degree similarity, which considerably complicates target recognition. Meanwhile, image is also obscured by background noise. To address these issues, this paper proposes novel oriented method named SPG-OSD. First, we propose an two-stage module scattering characteristics. Second, reduce false alarms missing ships, improve network incorporating characteristics first stage detector. A scattering-point-guided region proposal (RPN) designed predict possible key points make regression classification stages RPN increase vicinity Third, supervised contrastive introduced alleviate problem minute discrepancies among object classes. Region-of-Interest (RoI) loss proposed enhance inter-class distinction diminish intra-class variance. Extensive experiments are conducted dataset from Gaofen-3 satellite, experimental results demonstrate effectiveness SPG-OSD show that our achieves state-of-the-art performance.