EEMEFN: Low-Light Image Enhancement via Edge-Enhanced Multi-Exposure Fusion Network

Introduction

This project page provides TensorFlow 1.X code that implements the following AAAI2019 paper:

Title: "EEMEFN: Low-Light Image Enhancement via Edge-Enhanced Multi-Exposure Fusion Network"

Paper: https://aaai.org/ojs/index.php/AAAI/article/view/7013/6867

How to use

conda env create --prefix ./env --file environment.yml
conda activate .env
pip install h5py
pip install rawpy

Download pretrained model

Sony and Fuji model

# model directory structure
result
|-- Sony_edge
|-- Sony_merge
|-- Fuji_edge
└-- Fuji_merge

Process data

Please download raw dataset from https://github.com/cchen156/Learning-to-See-in-the-Dark
And delete misalignment images
Then process data as follows (about two hours)

python dataset.py

# dataset directory structure
dataset
|-- Fuji
|   |-- fuji_long.hdf5
|   |-- fuji_short.hdf5
|   |-- long
|   └-- short
└-- Sony
    |-- sony_long.hdf5
    |-- sony_short.hdf5
    |-- long
    └-- short

Training

Please modify the name of target dataset first

python train_fusion.py
python train_merge.py

Evaluation

python test.py

Performance

Model	Sony (PSNR)	Sony (SSIM)	Fuji (PSNR)	Fuji (SSIM)
baseline	29.06	0.787	26.95	0.717
MEF	29.43	0.791	27.21	0.719
EEMEFN (paper)	29.60	0.795	27.38	0.723
EEMEFN (pretrained model)	29.60	0.795	27.43	0.723

License

This code is released under the MIT License (refer to the LICENSE file for details).