Introduction

This is the implementation of the "parameter constrained spectral encoder and decoder (PCSED)" introduced in the paper "Deep-Learned Broadband Encoding Stochastic Filters for Computational Spectroscopic Instruments". Paper URL: https://doi.org/10.1002/adts.202000299

Instructions

Variables:

The names of the variables are not consistant with those concepts in the paper. Here are the matchups (the variable name in the left corresponds to the concept in paper in the right):
fnet -- FMN
hsnet -- SED
hybnet -- PCSED
rnet -- IDN (inverse design network)

Folders:

data -- dataset folder. We did not upload the dataset because it exceeds the Github repository storage limit. You can download the demonstration dataset from https://zenodo.org/record/5111584#.YPLiUj3iuAI or https://pan.baidu.com/s/12EBoRjEdTD7FQ4NIhzATJA (extraction code: 'best') and copy it to this folder, then the scripts should work correctly.
nets -- the folder for storaging the networks and the generated data.

Files:

HybridNet.py -- PCSED definition.
PSNR.py -- MSE to PSNR transformation function definitions.
run_fnet.py -- run the trained FMN.
run_hsnet.py -- run the trained SED.
run_hybnet.py -- run the trained PCSED.
show_HSI_error.py -- plot the hyperspectral image (HSI) reconstruction results of PCSED and SED.
train_fnet.py -- train an FMN.
train_hsnet.py -- train a SED.
train_hybnet_Meta.py -- train a PCSED.
train_tnet.py -- train an IDN using a tandem neural network architecture (this method is proposed in Ref. [18] of the main text).
redesign_hsnet.py -- re-design the ROFs of the SED. This script is for designing the ROFs using a different material (metasurface or thin-film) to match the same target spectral responses.