Low-light Image Enhancement with Domain Adaptation

In recent years, deep learning-based algorithms for low-light images have achieved SOTA performance. Traditional unsupervised learning methods typically use image-to-image transformation models to solve low-light image enhancement problems. However they do not perform well on real images and scenes because of the over-reliance on paired real and synthetic image datasets that do not suppress the noise in existing scenes. This is mainly due to domain migration. To address the effects of domain migration, we propose an unsupervised domain adaptation paradigm. This includes a different image translation module, an image illumination enhancement module and a noise suppression module. In order to reduce the impact of domain migration, we align two domains through a two-way translation network. We use the images before and after translation and join the network for training at the same time to maintain a consistency loss between the same group of images. At this stage, we incorporate real low-light images into the enhancement network training using the properties of well-illuminated images (Retinex theory) to further improve domain adaptation. Through ablation and comparative experiments, our proposed method outperforms recent approaches under various metrics about visual quality and subjective feelings.