SAR Image Despeckling Based on Differential Attention and Multiscale Deep Perception

Synthetic Aperture Radar (SAR) images are often degraded by multiplicative speckle noise, hindering their processing and analysis. While Convolutional Neural Networks (CNNs) have limited receptive fields, limiting global feature capture, Transformer-based methods struggle to differentiate between features and noise as network depth increases. We propose DAMSNet, a novel U-Net-based architecture designed for the efficient reduction of noise in SAR images while preserving fine details. Experiments on both synthetic and real SAR datasets show that DAMSNet has demonstrated superior performance over traditional methods and state-of-the-art deep learning algorithms. To the best of our knowledge, DAMSNet is the first to incorporate a differential attention mechanism for SAR image despeckling.