This is the code implementation of paper: "Hybrid Spatial and Frequency Network for Light Field Image Restoration"

• Pytorch == 2.3.0
• CUDA == 12.1

# git clone this repository
git clone https://github.com/duongvinh/HLFRN.git
cd HLFRN

Please first download light field datasets in here.

For sythesis test data

We provide the pre-trained models for adding zero-mean Gaussian noise with the standard variance varying in the range of 10, 20, and 50 on the Lytro dataset. Enter the scripts folder and run:

python test.py \
    --model_name HLFRN \
    --sigma 50 \
    --modelPath ./pretrained_models/HLFRN/model_sigma_50.pth \
    --dataPath  ./test_noiseLeve_10-20-50_4-11_5x5.mat \
    --savePath ./results/sythesis_img_test/ \

For real test data

We have used Lytro Illum camera to capture various real-world LF image under different conditions. Please download via Google Drive

python demo.py \
    --model_name HLFRN \
    --sigma 50 \
    --modelPath ./pretrained_models/HLFRN/model_sigma_50.pth \
    --dataPath ./data/ \
    --savePath ./results/real_img_test \

For test other methods, we just need to modify "--model_name" to MSP, DRLF, or PFE methods. More examples can be seen in "./scripts" folder.

Enter the scripts folder and run:

python train.py \
    --model_name HLFRN \
    --sigma 50 \
    --dataPath  ./train_noiseLevel_10-20-50_4-11_color_5x5.mat \
    --saveCheckpointsDir ./checkpoints/ \

For train other methods, we just need to modify "--model_name" to MSP, DRLF, or PFE methods. More examples can be seen in "./scripts" folder.

If our work is useful for your research, please consider citing:

@Article{vinh2023-lfsr,
  author  = {Duong, V. V. and Nguyen, T. H. and Yim, J. and Jeon, B.},
  journal = {submitted IEEE Trans. Compuational Imaging},
  title   = {Hybrid Spatial and Frequency Network for Light Field Image Restoration},
  year    = {2024},
}

We would like to thanks the authors of DRLF, PFE, and MSP for sharing code.

If you have any questions, please feel free to reach me out at duongvinh@skku.edu.