Abstract: A novel image denoising method based on a multi-scale dual-branch Transformer is proposed to address the limitations of existing algorithms, such as insufficient utilization of local features and the inability to effectively restore edge details, leading to distortions. Firstly, this method designs a dual-branch Transformer structure that effectively integrates multi-scale information from both shallow and deep features, allowing the preservation of global characteristics while restoring image details. Secondly, residual blocks are introduced to mitigate the potential gradient vanishing problem caused by network depth. Finally, a polarized self-attention mechanism is applied to enhance the model’s perception of multi-scale features, reducing the loss of feature information during down-sampling, while controlling the number of parameters. Experimental results demonstrate that the proposed method not only effectively removes noise but also restores finer texture details, outperforming existing mainstream denoising methods in both qualitative and quantitative evaluations.