Efficient single-image de-raining network based on self-attention and knowledge distillation

Single-image de-raining remains a challenging task in computer vision, particularly in complex rainy scenes where preserving crucial texture details is essential. To address this issue, we propose an efficient single-image de-raining network that integrates a multi-scale self-attention module and a lightweight knowledge distillation strategy. The self-attention module enhances the network’s ability to distinguish rain streaks from background textures by incorporating both channel and spatial attention mechanisms. Meanwhile, the knowledge distillation strategy reduces the model’s computational complexity while maintaining performance. Extensive experiments on the Rain100L and Rain100H datasets demonstrate that our method surpasses existing state-of-the-art approaches in both de-raining quality and computational efficiency.