MDnT: A Multi-Scale Denoising Transformer Beyond Real Noisy Image Denoising

Real noisy image denoising is a challenging low-level vision task that aims to remove noisy signals from degraded realworld images and recover original clean image. Recently, CNN-based and Transformer-based denoising researches have shown impressive results, however, most of the work neglect to capture the long-range dependencies, and the potential progressive relationships of hierarchical features. To address these issues simultaneously, we propose a multi-scale denoising Transformer (MDnT). Specifically, MDnT is mainly composed of stacked multi-scale residual Transformer blocks (MRTBs). MRTB enables the model to maintain spatial details from low-level features while capturing rich semantics from high-level features by constructing parallel multi-scale flows. Further, we introduce a multi-scale feature fusion layer (MFFL) to accomplish the information interaction between scales in order to enhance the richness of the learned representations. Extensive experiments are conducted on publicly real noise datasets for verification. The results show that our method achieves a competitive performance to state-of-the-art methods, while the total number of multiplyaccumulate operations can be reduced by up to 71%.