DSMF-Net: A One-Stage SAR Ship Detection Network Based on Deformable Strip Convolution and Multiscale Feature Refinement and Fusion

Synthetic Aperture Radar (SAR), an all-weather and day-and-night remote sensing imaging technology, is crucial for ship detection. However, SAR images are hampered by speckle noise and coastal clutter, and ship targets exhibit multi-scale and small-scale characteristics. To tackle these challenges, we introduce the DSMF-Net, a SAR ship detection network leveraging deformable strip convolution and multi-scale feature refinement and fusion. Firstly, to counter interference from complex backgrounds like nearshore ports and speckle noise, the deformable strip convolution (DSConv) is introduced and incorporated into the backbone network for SAR ship feature extraction, named SSFEBackbone. DSConv adaptively adjusts convolution sampling positions based on ship target characteristics, precisely extracting features with directional and strip structures. Secondly, the dual-stream self-attention feature refinement module (DSFRM) is utilized to refine high-level semantic features. Through the mixing spatial and channel attention mechanism (MSCA), differences and correlations between complex backgrounds and ship entities are further captured, enhancing feature expression. Finally, the adaptive selective feature pyramid network (AS-FPN) is proposed. By leveraging MSCA attention, high-level semantic and low-level spatial features are flexibly matched, enabling better key information retention during fusion and background clutter suppression, thus improving detection performance for complex backgrounds and multi-scale ship targets. Experimental results demonstrate that DSMF-Net performs significantly better in ship detection in SAR images. It outperforms existing state-of-the-art methods on the SAR-Ship-Dataset, high-resolution SAR images dataset (HRSID), and SAR ship detection dataset (SSDD), achieving an AP<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> of 96.8%, 93.1%, and 97.4%, respectively.