Single image deraining via a recurrent multi-attention enhancement network

Single image deraining has attracted considerable attention because rain streaks can severely degrade image quality and affect performance in computer vision tasks. While recurrent networks have achieved promising results, they still have limitations such as their receptive field size and feature representation ability, increasing the difficulty in removing long rain streaks while preserving details. In this paper, we present a recurrent multi-attention enhancement network for single image deraining that uses multiple attention mechanisms to effectively enhance feature representation in two stages. In the first stage, we utilize a non-local block to enhance the attention of the location information, effectively expanding the receptive field and helping to remove long rain streaks. In the second stage, a feature-enhanced residual block (FERB) is used to effectively acquire useful channel and spatial information to preserve more details. The proposed network utilizes the results of the current cycle in the next cycle and reuses the original rainy image at each cycle to produce more details in the derained image. Extensive experiments show that our method outperforms state-of-the-art methods in terms of visual effects and quantitative results.