iCycleGAN: An Improved CycleGAN for Rain Streak Removal From Single Image

On the one hand, although the supervised learning methods have been used for image rain removal task, such methods have obvious limitations because maybe there is no or only few paired images with-without rain. On the other hand, problems such as color distortion and the inpainting of background information is not clear enough also limit the processing effect of unsupervised methods for image rain removal. An improved CycleGAN (iCycleGAN) was proposed to remove rain streak from a single image. First of all, CycleGAN's transfer learning ability and cyclic structure can solve the problem of the lack of paired data sets. Secondly, a densely connected convolutional network (DenseNet) was added to the generator backbone network to improve the protection of high-frequency information such as background textures, and a CBAM attention mechanism was added to the generator to focus on the repaired area near the rain streak and obtain a clearer repaired image. Finally, feature perceptual loss was introduced to strengthen the constraint of image feature restoration and obtain more realistic results. In order to verify the effectiveness of the proposed method, training was conducted on Rain100L and Rian800 data sets. The comparison of experimental results shows that the model is superior to the existing unsupervised methods in the overall repair effect, and also has comparable inpainting effect compared with mainstream supervised methods.