Multi-scale network toward real-world image denoising

Images are inevitably degraded when captured due to the effects of noise, and thus denoising is required. Previous methods remove real-world noise, while also causing issues with over-smoothing image details and loss of edge information. To solve these issues, a multi-scale image denoising network (MSIDNet) is proposed in this paper. We design a residual attention block (RAB) to encode and decode the context well, while introducing a selective kernel feature fusion module to fuse multi-scale features and obtain rich contextual information from low-resolutions to restore more details. A feature extraction block (FEB) is designed to fully extract local and global features then fusion, which obtains rich feature information. Extensive experiments on four real-world image datasets demonstrate that our method has excellent generalization and achieves advanced denoising performance on both peak signal-to-noise ratio and structural similarity. MSIDNet preserves more edge details and improves the over-smoothing issue to enhance the visual effect of denoised images.