Unsupervised deraining with domain translation

Single-image rain removal is an essential preprocessing step, since rain streaks can significantly deteriorate the quality of an image and hinder subsequent outdoor multimedia applications. Existing deraining methods achieve excellent performance by using mass paired synthetic data, but their generalization ability is limited in practical applications. In this paper, we introduce a Domain-Aware Unidirectional Generative Adversarial Network (DAU-GAN) designed to capture desired characteristics in a unidirectional unsupervised manner for the task of single-image deraining. To improve the extraction of domain-relevant features in the generative process, we propose a novel attention-concatenated stacked U-Net module. This module obtains context-aware features across different scales through down- and up-sampling procedures. To effectively obtain derained images, we utilize a region-aware discriminator to differentiate between real and synthetic images. Extensive experiments demonstrate that our designed DAU-GAN outperforms state-of-the-art approaches on a range of frequently used datasets, particularly in real-world databases.