A Frequency Self-Attention MIMO Network for SAR Weak Targets Denoising and Sidelobe Suppression

Synthetic aperture radar (SAR) is affected by speckle noise and sidelobe artifact because its inherent imaging characteristics, which affect the application in the high-level task. Especially in the field of weak targets detection of SAR images, the intensity of target features is close to the intensity of speckle noise, which can lead to the denoising algorithm suppressing both the target textures and speckle noise. In order to preserving weak texture while denoising and suppressing sidelobe, we propose a HiLo-R edge attention network (HEANet) for SAR Weak Targets denoising and sidelobe suppression. HEANet uses multi-inputs and multi-outputs (MIMO) structure and the HiLo-R module extract diverse feature information from different domains. In addition, a predicted edge feature fusion (PEFF) is introduced to fuse features from different scales and enhance edge features, together with a sidelobe loss and frequency edge loss to improve edge feature preservation and sidelobes suppression. HEANet contains more frequency domain information and places greater emphasis on edge features, which is beneficial for better maintaining weak targets during the denoising process. The experimental results show that the proposed HEANet outperforms than state-of-the-art SAR image denoising methods, particularly in preserving weak detailed edge feature and has capable of suppressing sidelobes.