LoFi: Neural Local Fields for Scalable Image Reconstruction

This repository is the official Pytorch implementation of "LoFi: Neural Local Fields for Scalable Image Reconstruction" published in IEEE Transactions on Computational Imaging.

Colab demo

Requirements

(This code is tested with PyTorch 1.12.1, Python 3.8.3, CUDA 11.6 and cuDNN 7.)

• numpy
• scipy
• matplotlib
• imageio
• torch==1.12.1
• torchvision=0.13.1

Installation

Run the following code to install conda environment "LoFi.yml":

conda env create -f LoFi.yml

Dataset

All datasets have been uploaded to SwitchDrive. The datasets can be downloaded using the following commands:

LDCT dataset:

curl -O -J https://drive.switch.ch/index.php/s/z9iqtogRJhdoGS0/download

Galaxies dataset:

curl -O -J https://drive.switch.ch/index.php/s/J93jxcl7qxjMwmH/download

ten-pic dataset:

curl -O -J https://drive.switch.ch/index.php/s/jnxhJ0ztAX1QHvK/download

After downloading the datasets, please put them in datasets directory in the main LoFi folder.

Experiments

Training & Inference

All arguments for training are explained in 'config.py'. After specifying your arguments, you can run the following command to train the model:

python3 train.py 

Citation

If you find the code useful in your research, please consider citing the paper.

@article{khorashadizadeh2024lofi,
  title={Lofi: Neural local fields for scalable image reconstruction},
  author={Khorashadizadeh, AmirEhsan and Liaudat, Tob{\i}as I and Liu, Tianlin and McEwen, Jason D and Dokmanic, Ivan},
  journal={arXiv preprint arXiv},
  volume={2411},
  year={2024}
}