Unsupervised Low-Light Image Enhancement from a Single Raw Image

A raw image includes much more original information of a real-world scene than the corresponding sRGB image. This is crucial for low-light image enhancement (LLIE), especially for extreme low-light conditions. Existing supervised learning based LLIE algorithms convert a raw data of a short-exposed image into the corresponding long-exposure image in sRGB space. However, they have a fundamental limitation, i.e., they are not able to suppress over-exposed regions. In this paper, a novel unsupervised learning based LLIE algorithm is introduced to convert a raw data of a short-exposed image into an sRGB image which is fused from multiple sRGB images with different exposures around a target one. As such, the proposed LLIE algorithm overcomes the fundamental limitation of the existing LLIE algorithms. Experiment results validate the proposed LLIE algorithm.