Learning Contextual Embedding Deep Networks for Accurate and Efficient Image Deraining.

The existing state-of-the-art deraining methods rely on a bulky network structure to accurately capture rain streaks, resulting in prohibitive memory consumption and computation cost. In addition, most of them destroy background details along with the process of rain removal. This paper presents a simple but effective network to address the above problems. Firstly, we developed a lightweight model, called Cross-Scale Contextual Embedding Network (CSCEN), to achieve the cross-scale embedding of rain streaks in different scale contexts. We also introduced a new deeply-supervised rain detection loss, making the training process of the intermediate layers be direct and transparent thus the transfer of rain information layer-by-layer would be under appropriate supervision. Qualitative and quantitative experiments show that the proposed network is superior to the advanced deraining methods and inherits better generalization in real-world rain removal.