Image Brightness Adjustment with Unpaired Training

In camera-based imaging, exposure correction is an important issue. Wrong exposures include overexposure and underexposure, while prior works mainly focus on underexposed images or general image enhancement. However, it is challenging to collect images of different brightness levels in practice. This paper proposes a network structure based on Generative Adversarial Network (GAN) trained on unpaired image datasets. Instead of using ground truth data to supervise learning, we adopt information extracted from the input to regular unpaired training. Our method uses style transfer for reference to separate the content space from the brightness space, realizing the conversion of an image of arbitrary brightness into images of different brightness. A large number of experiments demonstrate that our method realizes the conversion of different brightness. Our work can restore the overexposed and underexposed images of different levels and enhance low-illuminance images effectively. We organize a dataset with different exposure levels based on the existing dataset as our training dataset.