Spatial frequency domain information aggregation network for radar image despeckling

Radar images are affected by speckle noise, which seriously affects its subsequent applications. Recently, the Convolutional Neural Networks (CNN) have been widely used in radar images despeckling and achieved remarkable results due to its powerful learning ability. However, most existing deep learning-based despeckling methods recover clear images in the spatial domain and rarely explore potential solutions in the frequency domain. This paper is the first attempt to combine spatial and frequency domain for radar image denoising, and proposes a spatial-frequency domain aggregation network for synthetic aperture radar (SAR) images despeckling, called as SFSARNet. SFSARNet contains three key components: a spatial-domain information branch, a frequency-domain information branch, and a dual-domain aggregation module. The spatial domain information branch mainly works in the spatial domain and is used to explore the local structural details. The frequency domain information branch works in the frequency domain for exploring global context information. The dual-domain aggregation module learns and integrates complementary information from both the spatial and frequency domains, enabling the network to benefit from both local and global context information. The proposed SFSARNet is validated on simulated and real radar image. Both quantitative and qualitative comparisons demonstrate that the proposed SFSARNet outperforms the existing mainstream methods.