DCTNet: Integrating Deformable Convolution and Transformer-Based Feature Fusion for High-Performance SAR Ship Detection

Synthetic aperture radar (SAR)–based ship detection has become an indispensable technique for applications in maritime surveillance, oceanic defense systems, and marine resource assessment. However, factors such as speckle noise, sea clutter, and the diversity of ship scales severely restrict the robustness and accuracy of detection. To address these challenges, this paper proposes a one-stage detection network, termed DCTNet, which integrates a Deformable Convolution-based Multi-scale Fusion (DCMF) module and a Transformer-enhanced Feature Pyramid Aggregation Network (TFPAN). The DCMF module employs deformable convolution and channel partitioning to adaptively capture multi-scale features, thereby enhancing the representation of small-scale or low-contrast ships while effectively suppressing background interference. The TFPAN module incorporates Transformer attention mechanisms with progressive feature aggregation to strengthen multi-scale feature interaction and semantic consistency. In addition, a Triple Feature Fusion (TFF) strategy is designed to preserve local details and ensure global context transmission, effectively mitigating the loss of small target features during the fusion process. Experimental results on the SAR-Ship-Dataset and HRSID demonstrate that the proposed DCTNet achieves excellent performance and exhibits strong generalization ability.