Dynamic Multi-Domain Translation Network For Single Image Deraining

On the single image deraining task, traditional methods are too complicated while deep learning methods lack interpretability. To solve these issues, we propose a novel deep unfolding network, which has the advantages of low complexity and high interpretability. Specifically, by transforming the rain into high-dimensional features, we propose to utilize the proximal gradient descend technique to construct an algorithm. And a new symmetry constraint is introduced to reduce the algorithm complexity effectively. Furthermore, to enhance the representation of rain features, we propose a novel dynamic multi-domain translation (DMT) module. Finally, by unrolling the algorithm, a deep unfolding network named DMTNet is established. All the parameters in DMTNet are learned end-to-end. Extensive experimental results show that the proposed DMTNet outperforms SOTA methods on several benchmark datasets.