Semi-supervised Single Image Deraining with Discrete Wavelet Transform

Abstract

In recent years, single image deraining has received considerable research interests. Supervised learning is widely adopted for training dedicated deraining networks to achieve promising results on synthetic datasets, while limiting in handling real-world rainy images. Unsupervised and semi-supervised learning-based deranining methods have been studied to improve the performance on real cases, but their quantitative results are still inferior. In this paper, we propose to address this crucial issue for image deraining in terms of backbone architecture and the strategy of semi-supervised learning. First, in terms of network architecture, we propose an attentive image deraining network (AIDNet), where residual attention block is proposed to exploit the beneficial deep feature from the rain streak layer to background image layer. Then, different from the traditional semi-supervised method by enforcing the consistency of rain pattern distribution between real rainy images and synthetic rainy images, we explore the correlation between the real clean images and the predicted background image by imposing adversarial losses in wavelet space $I _{HH}$, $I _{HL}$, and $I _{LH}$, resulting in the final AID-DWT model. Extensive experiments on both synthetic and real-world rainy images have validated that our AID-DWT can achieve better deraining results than not only existing semi-supervised deraining methods qualitatively but also outperform state-of-the-art supervised deraining methods quantitatively. All the source code and pre-trained models are available at https://github.com/cuiyixin555/DeRain-DWT.