Wavelet-Enhanced Multi-scale Swin-Deformable DETR for SAR Ship Detection

In recent years, Synthetic aperture radar (SAR), a transformative microwave imaging technology, plays a critical role in ship target detection within complex maritime environments. To address the challenges of clutter suppression and multi-scale feature representation in SAR images faced by traditional deep learning detectors, this paper proposes the Wavelet-Enhanced Multi-scale Swin-Deformable DETR model. The model replaces the CNN backbone of traditional Deformable DETR with Swin Transformer to capture finer multi-scale features through a hierarchical window attention mechanism. Additionally, wavelet convolution is introduced to decompose features into different frequency sub-bands, enhancing the model’s ability to distinguish ship targets from sea clutter. Experiments on the SSDD dataset show that the proposed method achieves an overall average precision (AP) of 0.423 and a medium-scale object AP_m of 0.527, significantly outperforming Deformable DETR baselines with different backbones. This validates its robustness in multi-scale ship detection and complex backgrounds.