A Dual-Branch SAR Object detection Framework Using Wavelet Decomposition and Attention-Guided Feature Fusion

Synthetic Aperture Radar (SAR) object detection presents unique challenges due to noise, clutter, and the need for both local and global feature extraction. Traditional single-branch detectors, whether purely CNN or Transformer based, often encounter feature competition, resulting in suboptimal performance. In this paper, we propose a dual-branch network comprising a CNN branch (e.g., ConvNeXt) and a Transformer branch (e.g., Swin Transformer) to tackle these issues. By separating local and global feature extraction, we harness the inherent strengths of each architecture and mitigate feature competition. A frequency-guided attention module (FGMA) then aggregates their complementary representations, leading to more accurate and robust detection outcomes. Experimental results on multiple benchmark SAR datasets demonstrate that our method consistently outperforms single-branch baselines in terms of both quantitative metrics (e.g., mean Average Precision) and qualitative visual analysis. This framework, leveraging prior knowledge of CNNs and Transformers, provides a promising direction for advancing SAR object detection.