NightDNet: A Semi-Supervised Nighttime Haze Removal Frame Work for Single Image

Because the radiation of the night scene is relatively complex, compared to the situation where the atmospheric light is globally consistent during the day, the night image dehazing is more challenging. Due to the complex atmospheric imaging conditions, the night haze image synthesis method is not mature enough, and the synthesized night haze images are easy to distinguish. Existing night dehazing methods use hand-designed prior knowledge, but prior knowledge is not very stable for real scenes and is prone to serious failures. In this work, we propose a weakly supervised image translation method. This method is a generative confrontation network combining haze atmosphere imaging model to translate the image from hazy to clear domain. The train and test data we use is all public and real night outdoor images, so the trained model can generate realistic nighttime images without haze.