Deep Multi-Illumination Fusion for Low-Light Image Enhancement

In recent years, improving the visual quality of low-light images has attracted tremendous attention. Most of the existing deep learning approaches estimate the single illumination and then obtain the enhanced result according to the Retinex theory. However, only estimating the single illumination limits the solution space of the enhanced result, causing the unideal performance, e.g., color distortion, details loss, etc. To overcome the issues, we design a new Deep Multi-Illumination Fusion (denoted as DMIF) network to effectively handle low-light image enhancement. Specifically, we first construct an illumination estimation module to generate multiple illuminations to enlarge the solution space. We fuse these illuminations and aggregate their advantages by an illumination fusion algorithm to produce a final illumination. Finally, the enhanced result is obtained according to the Retinex theory. Plenty of experiments are conducted to fully indicate our effectiveness and superiority against other state-of-the-art methods.