Skip to content

Latest commit

 

History

History

2021_Optics_express

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
.gitignore
.gitignore
 
 
README.md
README.md
 
 
main.ipynb
main.ipynb
 
 
main.py
main.py
 
 
train.py
train.py
 
 
train_them_all.ps1
train_them_all.ps1
 
 

README.md

Single-pixel Image Reconstruction from Experimental Data using Neural Networks

This repository contains the code that produces the results reported in

Antonio Lorente Mur, Pierre Leclerc, Françoise Peyrin, and Nicolas Ducros, "Single-pixel image reconstruction from experimental data using neural networks," Opt. Express 29, 17097-17110 (2021). DOI (open access).

Contact: [email protected], CREATIS Laboratory, University of Lyon, France.

Get the scripts

Clone the repo and navigate to the folder that corresponds to this study

git clone https://github.com/openspyrit/spyrit-examples.git
cd ./spyrit-examples/2021_Optics_express/

This folder contains two scripts

  • main.py: generates the figures in the paper (typically run in Spyder). It requires trained networks.
  • train.py: can be used to train the reconstruction networks (typically run in a terminal).

Install the dependencies

Our scripts primarily relies on the SPyRiT package (tested with v1.1.2) that can be installed via pip . We recommend creating a virtual (e.g., conda) environment first.

NB: On Windows, you need to install torch before SPyRiT.

# conda (or pip) install
conda create --name spyrit-env
conda activate spyrit-env
conda install -c anaconda spyder 
conda install -c anaconda pip
# for windows only
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

# pip install
pip install spyrit # tested with spyrit==1.1.2
pip install pylops

Experimental data

We exploit some of the single-pixel raw acquisitions that constitute the SPIHIM dataset. The preprocessed data considered in the paper can be downloaded here. Unzip in ./spyrit-examples/2021_Optics_express/expe/.

Trained neural networks

We provide the trained network here. Unzip in ./spyrit-examples/2021_Optics_express/model/.

Training the neural networks from scratch

Type of networks

We provide train.py to train several variants of networks, from a single command line. This includes

  • DC-Net: Denoised Completion (Tikhonov) network
python train.py
  • C-Net: Completion network (DC-Net with no denoising)
python train.py --no_denoi
  • Free-Net. This is a C-Net where the completion layer is learnt (i.e., free during training)
python train.py --no_denoi --net_arch 3
  • Noiseless-Net. This is a C-Net trained with no noise (i.e., infinite number of photons)
python train.py --no_denoi --intensity_max 'inf'

By default, all networks are trained for M = 512 measurements during 20 epochs. For other values, consider

python train.py --CR 1024 --num_epochs 100

To consider different noise levels, play with the maximum number of photons (default, 2500 photons). For instance

python train.py --intensity_max 50

Image database

We train our networks by simulating single-pixel measurements from the STL-10 image database. Running the above command lines will download the database.

[optional] If you already have the STL-10 on your computer, create a symbolic link.

  • Linux: [CHECK]
ln -s <stl-10 parent folder> /data/
  • Windows Powershell:
New-Item -ItemType SymbolicLink -Name \data\ -Target <stl-10 parent folder>