Learning-Based Reconstruction of FRI Signals

Vincent C. H. Leung (chi.leung14@imperial.ac.uk), Jun-Jie Huang, Yonina C. Eldar and Pier Luigi Dragotti

This repository is contains the code and pretrained models for "Learning-Based Reconstruction of FRI Signals" (TSP' 23; arxiv).

The code consists of two different learning-based FRI models: Deep Unfolded Projected Wirtinger Gradient Descent (Deep Unfolded PWGD) and FRI Encoder-Decoder Network (FRIED-Net). Deep Unfolded PWGD operates in the frequency domain and unfolds the iterative denoising process, while FRIED-Net models the FRI acquisition process and the fact that FRI signals are determined by a small number of parameters to form an autoencoder-like network architecture.

Prerequisites

The code was tested on Python 3.10 so this particular version is recommended.

The code for both models was developed based on PyTorch. torch>=2.0.0 is required specifically for Deep Unfolded PWGD since it deals with complex numbers. You can install the latest version of PyTorch here.

Other required packages are listed in requirements.txt.

Deep Unfolded Projected Wirtinger Gradient Descent (Deep Unfolded PWGD)

Please refer to DeepUnfoldedPWGD for the implementation and pretrained models.

FRI Encoder-Decoder Network (FRIED-Net)

Please refer to FRIED-Net for the implementation and pretrained models.

Dataset

Two types of data are used in the paper: synthetic and calcium imaging (cai-1) dataset. Please refer to datasets for related MATLAB codes and detailed instructions on how to use them.

Reference

If you find this repository, please cite the following paper:

V. C. H. Leung, J.-J. Huang, Y. C. Eldar and P. L. Dragotti. "Learning-Based Reconstruction of FRI signals," IEEE Transactions on Signal Processing, vol. 71, pp. 2564-2578, 2023.

@article{Leung2023,
  title = {Learning-Based Reconstruction of {{FRI}} Signals},
  author = {Leung, Vincent C. H. and Huang, Jun-Jie and Eldar, Yonina C. and Dragotti, Pier Luigi},
  date = {2023},
  journal={IEEE Transactions on Signal Processing},
  volume={71},
  number={},
  pages={2564--2578},
  doi = {10.1109/TSP.2023.3290355}
}