A deep learning denoising method based on pseudo labels for airborne SAR images

Synthetic Aperture Radar (SAR) imaging is widely used in fields such as resource monitoring and topographic mapping due to its high resolution,strong penetration and all-weather adaptability. As an interferometric measurement technique, SAR utilizes the coherent superposition of multiple reflected signals to enhance the cross-range resolution of images. However, the phase fluctuation of the reflected signal can bring signal-dependent noise, namely speckle noise. The denoising of SAR images has been a hot research focus in the field. The current SAR acquisition platforms mainly include spaceborne SAR and airborne SAR. Under airborne conditions, relative motion between objects and the platform is more obvious, leading to more significant positional changes of imaged objects between time frames in the same field of view. Therefore, the denoising method guided by generating low-noise images by averaging multiple time frames is not suitable. In this paper, we propose a multiple-expert model distillation framework for self-supervised denoising. We leverage the adaptive capacity of multiple experts while mitigating noise correlation between neighboring pixels in SAR images through downsampling. A consistency loss is incorporated to preserve the denoising performance of the original on the original,higher-resolution image. Experimental results demonstrate that this method achieves superior denoising results on airborne SAR images compared to representative methods DIP and ZS-N2N.