FCA-Net: Feature Collaborative Enhancement Network for SAR Ship Detection in Complex Scenes

Abstract

Ship detection in synthetic aperture radar (SAR) images remains challenging because ship targets are often embedded in cluttered backgrounds and exhibit substantial structural and scale variations. To address these issues, a feature collaborative enhancement network, termed FCA-Net, is proposed for SAR ship detection in complex scenes. First, a directional feature refinement (DFR) block is embedded into the backbone to enhance anisotropic structural representation with direction-aware local features. Second, a spectral-guided feature modulation (SGFM) block is introduced into the lateral connections between the backbone and the feature pyramid network to decouple high- and low-frequency responses and improve feature propagation in cluttered SAR scenes. Third, a cross-level alignment fusion (CLAF) module is deployed in the neck to refine shallow structural priors and align them with deeper semantic features. Experiments on SSDD, HRSID, and SAR-Ship-Dataset validate the effectiveness of FCA-Net across different SAR ship detection scenarios. FCA-Net achieves AP50 of 98.1%, 92.1%, and 94.6%, and AP50:95 of 73.3%, 68.9%, and 61.3% on SSDD, HRSID, and SAR-Ship-Dataset, respectively. Ablation and visualization analyses further attribute the performance gains to the complementary roles of DFR, SGFM, and CLAF in local structural discrimination, feature propagation, and cross-level consistency, respectively. Overall, FCA-Net offers an effective solution for improving SAR ship detection in complex scenes.