Over-Exposure Correction via Exposure and Scene Information Disentanglement.

Over-exposure correction is an important problem of great consequence to social media industries. In this paper, we propose a novel model to tackle this task. Considering that reasonable enhanced results can still vary in terms of exposure, we do not strictly enforce the model to generate identical results with ground-truth images. On the contrary, we train the network to recover the lost scene information according to the existing information of the over-exposure images and generate naturalness-preserved images. Experiments compared with several state-of-the-art methods show the superior performance of the proposed network. Besides, we also verify our hypothesis with ablation studies. Our source code is available at https://github.com/0x437968/overexposure-correction-dise .