Removing rain based on a cycle generative adversarial network

Single image de-raining is a difficult problem due to its inherent ill-posed nature. Many researchers have done some related work and attempt to introduce prior information to get a good rain removal result. In this paper, we propose a CycleGAN-based new deep convolutional neural network (CNN) to solve single image de-raining problem. Our idea mainly comes from cycle generative adversarial networks (GANs). CycleGAN was initially introduced for style transfer task. However, we found that the thought of dual learning used in CycleGAN is also applicable for rain streaks removal. The separation of the rain streak and the background is the most important and the most difficult part of this problem. By adding a reverse mapping, dual learning can thoroughly remove rain streaks with the background been preserved and ease the training of both generators. Inspired by the deep residual network (ResNet), we also introduce residual image learning to our generators, which directly reduce the mapping range from input to output and make the mapping process easier. Furthermore, In order to prevent the model from over fitting on the synthetic training data so that the real-world rain line cannot be effectively removed, we add the Squeeze-and-Excitation (SE) block after our generator's convolution layers, which can push the network to learn more useful features, reduce the calculation, and reinforce the network's generalization ability. Test results based on real and synthetic images show our method outperforms many recent state-of-the-art single image de-raining methods in visual performance and Quantitative aspects.