DnT: Learning Unsupervised Denoising Transformer from Single Noisy Image

Abstract

In the last few years, a myriad of Transformer based methods have drawn considerable attention due to their outstanding performance on various computer vision tasks. However, most image denoising methods are based on convolutional neural networks (CNNs), few attempts have been made with Transformer, especially in self-supervised and unsupervised methods. In this paper, we propose a novel and good performance unsupervised image Denoising Transformer (DnT) which is just trained by the single input noisy image. Our network combines Transformer and CNN to predict the counterpart clean target, the training loss was measured by pairs of noisy independent images constructed from the input image. The dropout-based ensemble is used to get the final denoised result by averaging multiple predictions generated by the trained model. Experiments show that the proposed method not only has superior performance over the state-of-the-art single noisy image denoiser on additive white Gaussian noise (AWGN) removal but also achieves good results on real-world image denoising.