Cycle-Derain: Enhanced CycleGAN for Single Image Deraining

Abstract

As the basis of image processing, Single Image Deraining (SID) has always been a significant and challenging theme. On the one hand, due to lack of enough real rainy images and corresponding clean images, most derain networks train in the synthetic datasets, which makes the outputs unsatisfactory in real environment. On the other hand, heavy rainfall is accompanied by fog. Traditional networks for deraining is used to remove the rain streaks in the rain image. The processed image may still have the problem of blurring. In this paper, we comprehensively consider the problems existing in SID, propose a Cycle-Derain network based on unsupervised attention mechanism. Specifically, this network makes full use of generative adversarial networks with two mappings and cycle consistency loss to train the unpaired rainy images and clean images. Besides, it introduces unsupervised attention mechanism and uses the loop-search positioning algorithm to make the network better deal with the details of rain and fog in images. Many experiments based on public datasets show that Cycle-Derain network is very competitive with other rain-removing networks, especially in the restoration of real rainy images.KeywordsImage derainingGenerative adversarial networkAttention mechanism