Learning a Residual Fusion Network for SAR Image Despeckling

Deep convolutional neural networks (CNN) have recently been proven to be valid for synthetic aperture radar (SAR) image despeckling. However, these CNN-based despeckling algorithms do not make the most of multi-level features, especially high-frequency features, therefore achieving relatively poor performance. We develop a novel residual fusion model for SAR image depeckling (SAR-RFN). We make the most of multi-level features of different convolutional layers. To be specific, we design residual fusion block (RFB) to extract rich local features through dilated convolutions. Local feature fusion is used to combine all features and learn more effective local features. Moreover, The amalgamation of local features from all RFBs into global features is undertaken through global feature fusion to achieve holistic learning. Extensive experimentation illustrates the supremacy of our SAR-RFN over current leading-edge approaches regarding both quantitative and qualitative measures.