Rain to Rain: Learning Real Rain Removal Without Ground Truth

Image deraining is a low-level restoration task that has become quite popular during the past decades. Although recent data-driven deraining models exhibit promising results, most of these models are trained on synthetic rain data sets which do not generalize well when applied to real rain images. While recent real-rain data sets have achieved favorable generalization performance, generating rain-free ground-truths can be tedious and time-consuming. To address this problem, in this work, we present rain to rain training, an unsupervised training method for single image deraining. Our experiments show that it is possible to train single image deraining models by using only rain images. This can be achieved by simply training models to map pairs of rain images. We also introduce the idea of using the least overlapping training pairs, a method of selecting adequate training pairs that enables rain to rain training to achieve equivalent deraining performance compared to supervised training.