SAR Image Despeckling via Efficient Multi-Scale Attention Enhanced U-Net
Zhenyu Guo, Weidong Hu, Jincheng Peng, Guozhen Hu, Minghao Feng, Ming Tao Zhou · 2025
Synthetic Aperture Radar (SAR) images are crucial for remote sensing and target recognition due to their all-weather, all-time imaging capabilities. However, speckle noise during imaging degrades image quality and affects high-level visual tasks. Traditional denoising methods (e.g., Lee, Frost, SAR-BM3D) struggle to balance noise suppression and structural detail preservation, while deep learning approaches (e.g., U-Net, Dncnn) face challenges in multi-scale feature fusion and attention design, causing computational redundancy and information loss. To address this, we propose a Multi-scale Efficient Attention U-Net (EMA-U-Net), utilizing parallel convolution kernels for multi-scale feature extraction and integrating cross-spatial learning with channel reshaping to enhance feature representation and structural preservation. Experiments show that EMA-U-Net outperforms state-of-the-art baselines in PSNR and SSIM, achieving both efficiency and accuracy, demonstrating the potential of multi-scale efficient attention for SAR image denoising.