Rain removal plays an important role in the restoration of degraded images. Recently, data-driven methods have achieved remarkable success. However, these approaches neglect that the appearance of rain is often accompanied by low light conditions, which will further degrade the image quality. Therefore, it is very indispensable to jointly remove the rain and enhance the light for real-world rain image restoration. In this paper, we aim to address this problem from two aspects. First, we proposed a novel entangled network, namely EMNet, which can remove the rain and enhance illumination in one go. Specifically, two encoder-decoder networks interact complementary information through entanglement structure, and parallel rain removal and illumination enhancement. Considering that the encoder-decoder structure is unreliable in preserving spatial details, we employ a detail recovery network to restore the desired fine texture. Second, we present a new synthetic dataset, namely DarkRain, to boost the development of rain image restoration algorithms in practical scenarios. DarkRain not only contains different degrees of rain, but also considers different lighting conditions, and more realistically simulates the rainfall in the real world. EMNet is extensively evaluated on the proposed benchmark and achieves state-of-the-art results. In addition, after a simple transformation, our method outshines existing methods in both rain removal and low-light image enhancement. The source code and dataset will be made publicly available later.
updated: Mon Aug 09 2021 08:46:15 GMT+0000 (UTC)
published: Mon Aug 09 2021 08:46:15 GMT+0000 (UTC)