DerainAttentionGAN: unsupervised single-image deraining using attention-guided generative adversarial networks

As an important task of computer vision, the single-image deraining (SID) methods tend to supervised learning in the previous research. However, most existing SID methods suffer from the inability to collect paired datasets needed by supervised learning in real scenarios. In this paper, we introduce a recent image translation model known as CycleGAN into SID and propose Derain Attention-Guide GAN (DerainAttentionGAN) that only requires unpaired datasets can effectively overcome the above limitation. The main work of this paper is as follows: We firstly inject an attention mechanism into the generator, which makes the rain-removing regions to be concentrated near the rain line to preserve background details. Secondly, a multiscale discriminator is used to discriminate the generated image from different scales to improve its quality. Finally, the perceptual-consistency loss and internal feature perceptual loss (interfeat loss) are introduced to reduce artificial features on the generated image and make it more realistic. Experiments results demonstrate that our work is superior to the current unsupervised learning methods in terms of both quantitative and qualitative, and have achieved comparable effects to other popular supervised learning methods.