Daniel Feijoo, Juan C. Benito, Alvaro Garcia, Marcos V. Conde

Abstract: Photography during night or in dark conditions typically suffers from noise, low light and blurring issues due to the dim environment and the common use of long exposure. Although Deblurring and Low-light Image Enhancement (LLIE) are related under these conditions, most approaches in image restoration solve these tasks separately. In this paper, we present an efficient and robust neural network for multi-task low-light image restoration. Instead of following the current tendency of Transformer-based models, we propose new attention mechanisms to enhance the receptive field of efficient CNNs. Our method reduces the computational costs in terms of parameters and MAC operations compared to previous methods. Our model, DarkIR, achieves new state-of-the-art results on the popular LOLBlur, LOLv2 and Real-LOLBlur datasets, being able to generalize on real-world night and dark images.

Low-light w/ blur	RetinexFormer	DarkIR (ours)
Low-light w/o blur	LEDNet	DarkIR (ours)

• Python == 3.10.12
• PyTorch == 2.5.1
• CUDA == 12.4
• Other required packages in requirements.txt

# git clone this repository
git clone /~https://github.com/Fundacion-Cidaut/DarkIR.git
cd DarkIR

# create python environment
python3 -m venv venv_DarkIR
source venv_DarkIR/bin/activate

# install python dependencies
pip install -r requirements.txt

The datasets used for training and/or evaluation are:

You can download each specific dataset and put it on the /data/datasets folder for testing.

We present results in different datasets for DarkIR of different sizes. While DarkIR-m has channel depth of 32, 3.31 M parameters and 7.25 GMACs, DarkIR-l has channel depth 64, 12.96 M parameters and 27.19 GMACs.

We present perceptual metrics for Real-LOLBlur dataset:

LOLBlur results were obtained training the network only in this dataset. Best results in LOLv2-real, LOLv2-synth and both LSRW were obtained in a multitask training of the three datasets with LOLBlur (getting 26.63 PSNR and 0.875 SSIM in this dataset). Finally Real-LOLBlur results were obtained with a model trained in LOLBlur.

In addition, we tested our DarkIR-m in Real-World LLIE unpaired Datasets (downloaded from Drive):

To check our results you could run the evaluation of DarkIR in each of the datasets:

• Download the weights of the model from OneDrive and put them in /models.
• run python testing.py -p ./options/test/<config.yml>. Default is LOLBlur.

You may also check the qualitative results in Real-LOLBlur and LLIE unpaired by running python testing_unpaired.py -p ./options/test/<config.yml>. Default is RealBlur.

You can restore a whole set of images in a folder by running:

python inference.py -i <folder_path>

Restored images will be saved in ./images/results.

To inference a video you can run

python inference_video.py -i /path/to/video.mp4

which will be saved in ./videos/results.

LOLv2-real

Low-light	SNR-Net	RetinexFormer	DarkIR (ours)	Ground Truth

LOLv2-synth

Low-light	SNR-Net	RetinexFormer	DarkIR (ours)	Ground Truth

Real-LOLBlur-Night

If you have any questions, please contact danfei@cidaut.es and marcos.conde@uni-wuerzburg.de