KinD

This is a Tensorflow implementation of KinD

The KinD++ is an improved version.

Kindling the Darkness: a Practical Low-light Image Enhancer. In ACMMM2019
Yonghua Zhang, Jiawan Zhang, Xiaojie Guo

Paper

Requirements

1. Python
2. Tensorflow >= 1.10.0
3. numpy, PIL

Test

First download the pre-trained checkpoints from BaiduNetdisk or google drive, then just run

python evaluate.py

Our pre-trained model has changed. Thus, the results have some difference with the report in our paper. However, you can adjust the illumination ratio to get better results.

Train

The original LOLdataset can be downloaded from here. We rearrange the original LOLdataset and add several all-zero images to improve the decomposition results and restoration results. The new dataset can be download from BaiduNetdisk or google drive. Save training pairs of LOL dataset under './LOLdataset/our485/' and save evaluating pairs under './LOLdataset/eval15/'. For training, just run

python decomposition_net_train.py
python adjustment_net_train.py
python reflectance_restoration_net_train.py

You can also evaluate on the LOLdataset, just run

python evaluate_LOLdataset.py