Unsupervised Low-light Image Enhancement with Generated Low-light Image Pairs

Low-light image enhancement aims to improve the perception of images captured under low-light conditions. Many previous unsupervised methods rely solely on information from a single image and multiple priors for enhancement. However, the information within a single image is limited, and designing suitable priors could be challenging. Although some existing methods have addressed this using paired low-light images, obtaining such images is also intricate. To tackle this problem, we first generate paired low-light images with consistent content, noise independence, and slightly different illumination from a single low-light image. Second, we propose an unsupervised low-light enhancement network based on the paired images. Leveraging consistent image content, we establish mutual constraints between the two images to achieve identical enhancement results. To accomplish this, the Retinex theory is employed to decompose the images into illumination and reflectance components, ensuring consistency in the reflectance components of the two images, which facilitates the preservation of image content and avoids artifacts and color deviations. Moreover, as the images exhibit independent noise, we adopt the Noise2Noise for noise removal, ensuring comprehensive denoising without affecting image details, which further improves the quality of the enhanced images. Extensive experiments demonstrate that our approach has superior denoising capability while ensuring enhancement performance, and achieves results comparable to state-of-the-art methods.