Home

Awesome

TreEnhance: A tree search method for low-light image enhancement

Official Repo of the paper "TreEnhance: A Tree Search Method For Low-Light Image Enhancement" published in Pattern Recognition Journal

Marco Cotogni and Claudio Cusano

arxiv Paper

TreEnhance is an automatic method for low-light image enhancement, capable of improving the quality of digital images. It combines tree search theory and deep reinforcement learning to produce high-quality enhanced images.

The method has two main phases during training:

Two different inference solutions are proposed for the enhancement of new images:

The method also has a guided search strategy that "reverses" the enhancement procedure applied to a given input image. Unlike other methods, TreEnhance does not pose any constraint on the image resolution and can be used in a variety of scenarios with minimal tuning.

The method was tested on two datasets: the Low-Light dataset and the Adobe Five-K dataset, and obtained good results both qualitatively and quantitatively.

<p align="center"> <img src="figs/tree.png" width="400" height="350"/> <br/> <img src="figs/opt.png" width="400" height="150" /> </p>

Requirements

python > 3.7, Pytorch, Torchvision, PIL, numpy

Datasets

Download the LOL and Five-K datasets.

LOL Datasets

Split the data into train and test folders in according to the original papers.

Five-K

Once the dataset has been downloaded, split the data in training (4000) and test(1000) using the files train1+2-list.txt and test-list.txt

Reproducing the Experiments

In order to reproduce our experiments run the command

python3 training.py

to train the model and

python3 evaluation.py

to test the model and generate the images.

Parameter Explanation

Results

Quantitative results

<p float="left"> <img src="figs/lol_t.png" height=150 width="auto"/> <img src="figs/five-k.png" height=170 width="auto"/> </p>

Qualitative results

<p float="left"> <img src="figs/grid.png" width="350" height="400" /> <img src="figs/lol.png" width="350" height="400" /> </p>

Reference

If you are considering using our code, or you want to cite our paper please use:

@article{cotogni2022treenhance,
  title={TreEnhance: A Tree Search Method For Low-Light Image Enhancement},
  author={Cotogni, Marco and Cusano, Claudio},
  journal={Pattern Recognition},
  pages={109249},
  year={2022},
  publisher={Elsevier}
}