TU-Former: A Hybrid U-Shaped Transformer Network for SAR Image Denoising

In order to obtain a better synthetic aperture radar (SAR) image noise suppression effect, we combine Convolutional Neural Network (CNN) and Transformer network to construct a U-shaped hybrid Transformer (TU-former) for SAR image denoising. The encoder of the TU-former network consists of convolution, residual and self-attentive mechanisms. To avoid large-scale training of the model, TU-former first uses convolutional structure and residual structure for shallow feature extraction of images. Subsequently, TU-former performs long-term dependencies by using the self-attention mechanism of the Transformer block to further extract the deep features of the image. Finally, TU-former sums the output of the decoder with the input noise image to obtain the final denoised image. Compared with the state-of-the-art SAR image denoising algorithm, the proposed algorithm not only improves in each objective index but also shows great advantages in the visual effect after denoising.