A SAR image despeckling method combining Swin Transformer and DN-FPN

The presence of speckle noise in synthetic aperture radar (SAR) images has been shown to have a detrimental effect on the quality of the images, as well as presenting challenges when it comes to their processing and interpretation. Consequently, the removal of speckle noise from SAR images is a significant concern in the realm of remote sensing image processing. The transformer model, with its distinctive multi-head self-attention mechanism, has been demonstrated to excel in learning and modelling long-range dependencies in features. This capability surpasses that of traditional convolutional neural network (CNN)-based despeckling network, particularly in preserving image details. In this paper, we proposed a transformer-based despeckling network, which adopts the U-net network structure and introduces Swin transformer as an encoder to facilitate the extraction of image features with greater efficiency and accuracy. Furthermore, this paper introduces a DN-FPN design for feature-level denoising, which is a module that can reduce feature corruption and feature loss due to network passing or up-sampling during the training process. The experimental results demonstrate the efficacy of this approach, outperforming traditional CNN-based denoising methods on both simulated and real SAR images.