John-Venti/Self-DACE

Official pytorch version for Self-Reference Deep Adaptive Curve Estimation for Low-Light Image Enhancement

Self-DACE LLIE Method

Official pytorch version for Self-Reference Deep Adaptive Curve Estimation for Low-Light Image Enhancement

• Paper is avalible at:
• Old version: arXiv version - Self-Reference Deep Adaptive Curve Estimation for Low-Light Image Enhancement
• New version: will come up soon.

Hope AI would illuminate our unknown and invisible path to the future as it illuminates low-light images!

Demo

Demo on Low-light Image Enhancement

Visual comparison with original low-light image on LOL and SCIE dataset. The enhanced images of our method are on the top-right corners, and the input low-light images are on the bottom-left corners.

Demo on the Improvement of Low-light Face detection (New Version)

Demostration of improvement for Dark Face Detection task (CVPR UG2+ Challenge 2021) on DarkFace Dataset using Retinaface. The number on the top of box is the confidence score given by Retinaface with a confidence threshold of 0.5.

IoU=0.25	IoU=0.50	IoU=0.75

Test on the first 200 images.

Demo on the Improvement of Low-light Image Interactive Segmentation (Old Version)

Demostration of improvement for segmentation task on DarkFace Dataset using PiClick. The green stars are the objects of interactive segmentation what we want to segment. GT is annotated on the enhanced images manually by us.

New version Framework

Quantitative Comparison

Old Version: Ours* is the result only from Stage-I.

New Version:

Note:

• Our metric values are in line with Low-Light Image and Video Enhancement Using Deep Learning: A Survey.

Visual Comparison on LIME (Old version)

Visual Comparison on SCIE (New version)

How to use it

Prerequisite

cd ./codes_SelfDACE
pip install -r ./requirements.txt

Test Stage-I (only enhancing luminance)

cd ./stage1
python test_1stage.py

Test dates should be placed in codes_SelfDACE/stage1/data/test_data/low_eval, And then results would be found in codes_SelfDACE/stage1/data/result/low_eval.

Test both Stage-I and Stage-II (enhancing luminance and denoising)

cd ./stage2
python test_1stage.py

Test dates should be placed in codes_SelfDACE/stage2/data/test_data/low_eval, And then results would be found in codes_SelfDACE/stage2/data/result/low_eval.

How to train it

Prerequisite

cd ./codes_SelfDACE
pip install -r ./requirements.txt

Train Stage-I (only enhancing luminance)

1. You should download the training dataset from SCIE_part1 and resize all images to 256x256. Or you could download it directly from SCIE_part1_ZeroDCE_version, of which iamges have been cropped to 512x512 already. If you want to use it in your work, please cite SCIE_part1.

2. cd ./stage1
   python train_1stage.py
   

Train Stage-II (only denoising)

1. Copy the pre-trained model and training dataset from stage1, and put pre-trained model of Stage-I in ./stage2/snapshots_light

2. cd ./stage2
   python train_2stage.py
   

Acknowledgment

This paper gets a big inspiration from ZeroDCE.

Citation

If you find our work useful for your research, please cite our paper

@article{wen2023self,
  title={Self-Reference Deep Adaptive Curve Estimation for Low-Light Image Enhancement},
  author={Wen, Jianyu and Wu, Chenhao and Zhang, Tong and Yu, Yixuan and Swierczynski, Piotr},
  journal={arXiv preprint arXiv:2308.08197},
  year={2023}
}

• Thanks for all related work and workers.