Low Light Face Image Enhancement

For advanced CV project in University of Houston. Goals:
1) reimplement the state-of-art paper “Deep Retinex Decomposition for Low-Light Enhancement”.
2) test it on acquired low light faces.
3) run a pretrained face&landmark detection model to test improvement on low light faces.

Here is the paper website: https://daooshee.github.io/BMVC2018website/
The author also has published their code and dataset. You can find it on above website.

Documents

Final presentation
Final report
lowlight.ipynb: implementation of Retinex-Net
Traditional_enhance.ipynb: implemenation of Gamma enahncement
landmark.ipynb: face and landmark deteciton

Dependency

This work is done by tensorflow 1.10 + Python3.6.6
Training can be finished in 1 hour, with a GTX1060(6G RAM).

Network architecture

Result for image decomposition

{low Light, normal light, reflectance, illumination}

Result for image enhancement