MemDNet: Memorizing More Exogenous Information to Dehaze Natural Hazy Image

Model trained on synthetic datasets are difficult to obtain good results in the real world due to domain shift. Previous works have been devoted to mining the potential of existing natural datasets to obtain better results, but the number of images in the dataset is ultimately limited. In this paper, we propose MemDNet, a two-branch dehazing network that incorporates an atmospheric scattering model. Our approach utilizes a residual network and a U-shaped network to estimate the corresponding parameters. To alleviate the issue of insufficient natural hazy image pairs, we propose a generalized memory branch that incorporates information from exogenous images. This allows the network to memorize the fine details of the exogenous images, thereby enhancing its ability to recover image colors. Extensive experimental results demonstrate that our proposed memory branch is general and effective, and our method outperforms the state-of-the-art methods on real datasets.