Continuous Exposure for Extreme Low-Light Imaging

We consider the problem of enhancing an underexposed dark image captured in a very low-light environment where details cannot be detected. Existing methods learn to adjust the input image's exposure to a predetermined value. In practice, however, the optimal enhanced exposure varies from one input image to another, and as a result, the enhanced images may contain visual artifacts such as low-contrast or dark areas. We address this limitation by introducing a deep learning model that allows the user to continuously adjust the enhanced exposure level during runtime in order to optimize the output based on his preferences. We present a dataset of 1500 raw images captured in both outdoor and indoor scenes in extreme low-light conditions, with five different exposure levels and various camera parameters, as a key contribution. We demonstrate that, when compared to previous methods, our method can significantly improve the enhancement quality of images captured in extreme low-light conditions under a variety of conditions.