Single Image Deraining Network with Fusion Normalization and Rain-Streak Localization

In recent years, Transformers have surpassed CNNs in performance, becoming the SOTA network backbone. However, their computational complexity scales quadratically with image resolution, leading to lower efficiency in deraining task. In this paper, we build upon the efficient CNN-based NAFNet and design some key components, such as Fusion Normalization Block (FNB) and Auxiliary Rain-Streak Localization Branch Network (ARLBN), to form the Single Image Deraining Network with Fusion Normalization and Rain-Streak Localization (FNR-Net). It achieves a dual advantage of performance and efficiency. In single image deraining, FNR-Net performs comparably to the Transformer-based Restormer in terms of performance while significantly surpassing it in efficiency by 65%.