wangliwen1994/dln

low-light image enhancement

DLN (Lighting Network for Low-Light Image Enhancement)

By Li-Wen Wang, Zhi-Song Liu, Wan-Chi Siu, and Daniel P. K. Lun

This repo provides simple testing codes, pretrained models and the network strategy demo.

We propose a single image low-light enhancement method based on back-projection theory and attention mechanism. to achieve good enhancing performance.

BibTex

@ARTICLE{DLN2020,
  author={Li-Wen Wang and Zhi-Song Liu and Wan-Chi Siu and Daniel P.K. Lun},
  journal={IEEE Transactions on Image Processing}, 
  title={Lightening Network for Low-light Image Enhancement}, 
  year={2020},
  doi={10.1109/TIP.2020.3008396},
}

Complete Architecture

The complete architecture of Deep Lighten Network (DLN) is shown as follows, The rectangles and cubes denote the operations and feature maps respectively.

#Implementation

Prerequisites

• Python 3.5
• NVIDIA GPU + CUDA
• [optional] sacred+ mongodb (experiment control)

Getting Started

Installation

• Install PyTorch and dependencies from http://pytorch.org
• Install python libraries:

pip install pillow, opencv-python, scikit-image, sacred, pymongo

• Clone this repo

Testing

• A few example test images are included in the ./test_img folder.
• Please download trained model
  • Pretrained (trained at voc syntesised dataset that is more general) model from here (OneDrive link)
  • Fine-tuned at LOL dataset (towards real low-light image enhancement) from here (OneDrive link)
  • Put them under ./models/
• Test the model by:

python test.py --modelfile models/DLN_pretrained.pth

# or if the task towards real low-light image enhancement 
python test.py --modelfile models/DLN_finetune_LOL.pth

The test results will be saved to the folder: ./output.

Dataset

• Download the VOC2007 dataset and put it to "datasets/VOC2007/".
• Download the LOL dataset and put it to "datasets/LOL".

Training

It needs to manually switch the training dataset:

1. first, train from the synthesized dataset,
2. then, load the pretrained model and train from the real dataset

python train.py 

Quantitative Comparison

We tested the proposed method on the LOL real dataset for evaluation. We have achieve better performance.

Visual Comparison

At LOL dataset:

dln's Issues

about code

Can you explain the x = (x_ori - 0.5) * 2 in DLN code？Meanwhile,I can't find the interactive factor γ in the code. Is this factor necessary？The y = y + x in FusionLayer code didn't show up in the Fig.6 . I look forward to your reply.

about test set of synthetic images

Can you tell me which 100 images as test set of synthetic images?

Results on LOL dataset

Many thanks for your work and codes. I tested your model on LOL dataset by using the DLN_finetune_LOL.pth and got 21.946db which is the same as the result reported on your paper. But when I retrained the model, the result is only 19.45db. It seems there are some problems. So would you please give me some advise about this problem?
Many thanks.