MDDS-Net: Dual-Backbone Network with Dynamic Enhancement for SAR Ship Detection in Complex Maritime Environments

Ship detection in synthetic aperture radar (SAR) images is a key technology for maritime surveillance. However, practical applications face severe physical challenges, including low signal-to-clutter and signal-to-noise ratios under high sea states, discontinuous scattering of weak targets, and multitarget occlusion in dense ports. Existing methods struggle to suppress these severe interferences while preserving weak target structures. To address these issues, this article proposes a detection framework based on a dual-backbone complementary coordination mechanism. The method employs a parallel-designed local and global encoder to extract fine-grained spatial details of ships and model long-range scene dependencies, respectively. To explicitly tackle the aforementioned physical degradations, two core modules are introduced. First, a geometric prior-guided multiscale feature enhancement module integrates ship shape priors into spatial attention to aggregate discrete scattering centers, addressing weak target discontinuity and occlusion. Second, a bidirectional gated feature fusion module enables adaptive cross-backbone feature decoupling to suppress severe sea clutter in low SNR/SCR environments. In addition, an adaptive tiling strategy guided by land-sea masks is employed to efficiently process large-scale SAR images. Experiments demonstrate that multiscale dual-branch dependency synergy network achieves superior performance on the HRSID and RSDD-SAR datasets. Cross-validation using GF-3 satellite data and automatic identification system data further validates the robustness and practical value of our approach under complex maritime conditions.