Self Refining Deep Symmetry Enhanced Network for Rain Removal


Rain removal aims to remove the rain streaks on rain images. The state-of-the-art methods are mostly based on Con-volutional Neural Network (CNN). However, as CNN is notequivariant to object rotation, these methods are unsuitablefor dealing with the tilted rain streaks. To tackle this problem,we propose Deep Symmetry Enhanced Network (DSEN) thatis able to explicitly extract the rotation equivariant featuresfrom rain images. In addition, we design a self-refining mech-anism to remove the accumulated rain streaks in a coarse-to-fine manner. This mechanism reuses DSEN with a novel in-formation link which passes the gradient flow to the higherstages. Extensive experiments on both synthetic and real-world rain images show that our self-refining DSEN yieldsthe top performance.

International Conference on Image Processing