Single Image Deraining with Generative Adversarial Network

Abstract

Single image de-raining is an important computer vision task, which has attracted much attention from researchers in recently years. However, most current methods are still weak in image detail recovery. In this paper, we propose a rain removal algorithm based on the generative adversarial network (GAN) to better recover detailed information effectively. The generator utilizes U-net architecture along with a feature conversion module (FCM) to preserve image details. Additionally, we utilize high-level semantic loss, which uses the pre-trained VGG19 network to extract features between de-rained image and real image. The discriminator uses symmetric padding on each side of the feature maps to enlarge the receptive fields and even-sized kernels with little computational cost. We validate the effectiveness of our deraining algorithm through quantitative and qualitative analyses on Rain800, Rain200H and Rain200L datasets.