Exploiting Illumination Knowledge in the Real World for Low-Light Image Enhancement

Recently, deep-learning-based low-light image enhancement (LLIE) methods achieved promising results with large collections of data. However, low-/normal-light image pairs are difficult to obtain in practice. Although the indoor training pairs may be synthesized by manually adjusting the light exposure, there is currently no outdoor LLIE benchmark captured in the real world. To bridge the gap and benchmark outdoor LLIE tasks, we propose the first outdoor, real-world, 2-D, low-light dataset, dubbed RE2L. Based on RE2L, we propose a semisupervised LLIE framework to further exploit the illumination knowledge from both the signal fidelity constraint and characteristics of normal-light natural images. As the images in our RE2L dataset have various illumination conditions, we propose integrating the illumination degree information into the generator as guidance for controllable LLIE. Experimental results demonstrate that using RE2L for training deep LLIE schemes can improve the model effectiveness, both quantitatively and visually, especially on semisupervised LLIE.