FSCPUnet: A Blind Denoising U-shaped Network Based on Swin Transformer and Convolution
Maosen Ding, Ying Chen · 2023
Image denoising is a challenging and practical task that has long fascinated computer vision researchers. In past denoising efforts, Convolutional Neural Networks (CNNs) have been the most commonly used network model. In recent years, however, transformers have achieved considerable success in various visual tasks, including image restoration. Nevertheless, transformers still have limitations in extracting local image features compared to CNNs. Therefore, we aim to effectively combine Transformers and CNNs, harnessing their respective global and local feature extraction capabilities by proposing a novel image feature extraction module called SCPB and embedding it in a U-shaped network to enhance the multi-scale feature extraction capabilities. This gives rise to the FSCPUNet (Flexible Swin Transformer CAConv Parallel Unet) for image blind denoising. Furthermore, the model incorporates a noise estimation network to adapt to different noise levels, addressing the challenge of image blind denoising. Experimental results show that the FSCPUNet proposed in this paper outperforms existing denoising models in various denoising tasks.