A Novel Star Feature Decoupling Network for Multiscale SAR Image Ship Detection

Persistent challenges such as complex background clutter, low signal-to-noise ratios, and the frequent omission of small targets continue to hinder target detection performance in synthetic aperture radar (SAR) image. Compared to horizontal bounding box, oriented bounding box has lower computational efficiency due to more parameters optimization, although it has higher detection accuracy sometimes. Considering the tradeoff between computational efficiency and detection accuracy, the methods based on horizontal boundary box are still a choice which cannot be neglected in real applications. Recent advances in deep learning methods based on horizontal boundary box, particularly region-based and single-shot detectors, have improved accuracy through adaptive feature representation, yet issues like dense target scenes, cluttered near-shore backgrounds, and weak small-target features remain far from solved. To address these limitations, this article proposes a novel star feature decoupling network (SFD-Net) for multiscale SAR image ship detection. It includes two core contributions: a Feature Decoupled Mapping Module (FDMM), which adaptively separates discriminative and interfering features via a learnable subspace decomposition mechanism to enhance target-background contrast; and a star-fusion feature pyramid network (Star-Fusion FPN), which implicitly expands feature dimensions via elementwise multiplicative fusion to preserve critical fine-grained details of small targets. Extensive evaluations demonstrate the effectiveness of our method: on the SSDD dataset, SFD-Net achieves an Average Precision (AP) of 74.1% and an AP50 of 98.5% across all ship scales; on the HRSID dataset, it attains AP results of 70.9% and 92.5%, respectively. The results confirm consistent superiority over existing techniques and offer a robust feature learning strategy for SAR target detection under strong clutter interference.