A neural network approach to remove rain using reconstruction and feature losses

Rain streaks can eclipse some information of an image taken during rainfall which can degrade the performance of a vision system. While existing rain removing methods can recover the semantic structure, they lack natural texture recovery. The aim of this work is to recover the hidden structure and texture under the rain streaks with fine details. We propose a novel generative adversarial network with two discriminators to remove rain called rain removal generative adversarial network, where a combination of reconstruction, feature and adversarial losses is used for low level, structural and natural recovery, respectively. We have found that exploiting low-level loss with high-level structural similarity loss as a reconstruction loss is quite effective in attaining visually plausible and consistent texture. Qualitative and quantitative evaluations on our synthetically created dataset and a benchmark dataset show substantial performance gain than state-of-the-art rain removing methods.