Decompress Snapshot Compressive Imaging (DeSCI)

This repository contains the MATLAB code for the paper Rank Minimization for Snapshot Compressive Imaging (IEEE Transactions on Pattern Analysis and Machine Intelligence 2019) by Yang Liu*, Xin Yuan*, Jinli Suo, David J. Brady, and Qionghai Dai (*Equal contributions). [pdf] [github] [arXiv] [doi] [data (Google Drive)] [data (Baidu Drive)]

[New] Real data and associated code are available at this link on Google Drive or Baidu Drive. Note that the code for real data is not tested but with raw results as in the paper. Please refer to the readme file for the original source(s) of the real data.

Figure 1. Reconstructed Kobe video using DeSCI compared with the state-of-the-art methods, i.e., GMM-TP (TIP'14), MMLE-GMM (TIP'15), MMLE-MFA (TIP'15), and GAP-TV (ICIP'16). Here, 8 video frames are encoded in a single measurement and a total of 32 frames are presented by reconstructing 4 snapshot measurements. The Kobe video is used in the MMLE-GMM paper.

Figure 2. Reconstructed spectra of toy hyperspectral images using DeSCI compared with GAP-TV (ICIP'16). Here, 32 spectral frames are encoded in a single measurement. The toy hyperspectral images are from the CAVE multispectral image database.

Snapshot compressive imaging (SCI)

Snapshot compressive imaging (SCI) refers to encoding the three- or higher- dimensional data in a snapshot with distinct mask (or coded aperture) for each slice of the data. Decompress snapshot compressive imaging (DeSCI) then exploits the nonlocal self-similarity of natural scenes and applys an alternating minimization algorithm to solve the ill-posed problem.

Performance boost compared with the state-of-the-art methods includes more than 4 dB in terms of PSNR for simulated data and significant improvement for real data, which is addressed in the DeSCI paper. A video comparison of the proposed DeSCI method with the state-of-the-art algorithms is shown in Figure 1.

This code contains the simulated high-speed video Kobe dataset encoded with shifting masks following the coded aperture compressive temporal imager (CACTI), and the simulated hyperspectral toy dataset following the coded aperture shapshot spectral imager (CASSI). A brief review of the SCI systems, including CASSI and CACTI, and compressive light-field imaging is shown in Section 2 of the DeSCI paper. DeSCI could also be adapted for other compressive imaging systems with minor modifications, since we only need to change the sensing matrix for various coding strategies and the nonlocal self-similarity always hold for natural scenes. Code for the CACTI and CASSI data from real systems are available upon request.

Usage

Download the DeSCI repository

0. Requirements are MATLAB(R) with Parallel Computing Toolbox (parfor for multi-CPU acceleration).
1. Download this repository via git

git clone https://github.com/liuyang12/DeSCI

or download the zip file manually.

Run DeSCI on high-speed video

Kobe video data

2. Test the DeSCI algorithm (for high-speed imaging, that is CACTI on Kobe dataset) via

test_desci.m

and (optionally) video demonstrate the reconstruction results (after running test_desci.m) via

./figures/fig_desci_video.m

Run DeSCI on hyperspectral images

toy hyperspectral data

3. Test the DeSCI algorithm (for hyperspectral imaging, that is CASSI on toy dataset) via

test_desci_cassi.m

and (optionally) video demonstrate the reconstruction results (after running test_desci_cassi.m) via

./figures/fig_desci_cassi.m

bird hyperspectral data

[Optional] Test the DeSCI algorithm (for hyperspectral imaging, that is CASSI on bird dataset) via

./figures/test_desci_cassi_bird.m

and (optionally) video demonstrate the reconstruction results (after running ./figures/test_desci_cassi_bird.m) via

./figures/fig_desci_cassi_bird.m

Note that running the bird dataset is time- and memory- consuming for the size of the images is $1021\times 703$ with 24 spectral bands. Please ensure that the machine has at least 8 physical CPU cores and 32 GB memory to run ./figures/test_desci_cassi_bird.m, where more CPU cores and memory help for faster reconstruction.

Structure of directories

directory	description
algorithms	MATLAB functions of main algorithms proposed in the paper (original)
figures	MATLAB scripts to reproduce the results and figures in the paper (original)
packages	algorithms adapted from the state-of-art algorithms (adapted)
dataset	data used for reconstruction (simulated)
results	results of reconstruction (after reconstruction)
utils	utility functions

Platform

The test platform is MATLAB(R) 2017b (and 2018b) operating on Ubuntu 16.04 LTS (x64) with an Intel(R) Core(TM) 18-core processor at 2.60 GHz and 128 GB RAM. It can run on any machine with MATLAB(R) and Parallel Computing Toolbox, operating on Windows(R), Linux, or Mac OS. No GPU is needed to run this code.

It could take hours to run a single measurement depending on the number of frames collapsed into a single measurement and the number of CPU cores of the machine. This is due to the time-consuming iterations of block matching and low-rank approximation via singular value decomposition for weighted nuclear norm minimization of recovered video/spectral patches. The computation issue might be addressed by using generative adversarial networks (GAN) for block matching, and truncated SVD for low-rank estimation. Notice: Please donot wait for the results immediately after getting the DeSCI code to run.

Citation

@article{Liu19DeSCI,
   author  = {Liu, Yang and Yuan, Xin and Suo, Jinli and Brady, David J. and Dai, Qionghai},
   title   = {Rank Minimization for Snapshot Compressive Imaging},
   journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
   doi     = {10.1109/TPAMI.2018.2873587},
   year    = {2019},
   volume  = {41},
   number  = {12},
   pages   = {2990 - 3006},
   url     = {https://doi.org/10.1109/TPAMI.2018.2873587},
   type    = {Journal Article}
}

Contact

Yang Liu, Tsinghua University

Xin Yuan, Bell Labs