Dark-ControlNet: an enhanced dehazing universal plug-in based on the dark channel prior

Abstract

Existing dehazing models have excellent performance in synthetic scenes but still face the challenge of low robustness in real scenes. In this paper, we propose Dark-ControlNet, a generalized and enhanced dehazing plug-in that uses the dark channel prior as a control condition, which can be deployed on existing dehazing models and can be simply fine-tuned to enhance their robustness in real scenes while improving their dehazing performance. We first freeze the backbone network to preserve its encoding and decoding capabilities and input the dark channel prior with high robustness as conditional information to the plug-in network to obtain prior knowledge. Then, we fuse the dark channel prior features into the backbone network in the form of mean-variance alignment via the Haze&Dark(HD) module and guide the backbone network to decode clear images by fine-tuning the plug-in network. The experimental results show that the existing dehazing model enhanced by Dark-ControlNet performs well on synthetic datasets and real datasets.