Learning to See Faces in the Dark

Recovering details from dark images has received increasing attention due to its potential in applications such as video surveillance. We propose the first approach to detect and enhance human faces in extreme low-light images. Our method consists of two stages: a novel Face Location Network (FLNet) to locate the face, followed by a Face Enhancement Network (FE-Net) that uses concatenated sub-modules to progressively recover the face from coarse to fine grained details.
Specifically, our enhancement modules exploit the semantic priors of facial landmarks to facilitate face recovery. Extensive experiments show our method is quantitatively and qualitatively superior to the state-of-the-art in terms of enhancement quality and face recognition. We have also collected a real-world dataset to support relevant research. All code and data will be shared for reproducing our experiments.