Out-of-box Hyperspectral Image Restoration Toolbox
Denoising for remotely sensed images from QRNN3D
pip install hsir
Here are some runable examples, please refer to the code for more options.
python hsirun/train.py -a qrnn3d.qrnn3d
python hsirun/test.py -a qrnn3d.qrnn3d -r qrnn3d.pth -t icvl_512_50
Pretrained Models | Training Log | Datasets
Baidu Drive's Share Code=HSIR
Supported Models
Denoising | Super Resolution | Compressive Sensing | Spectral Reconstruction |
Gaussian Denoising on ICVL
Sigma=30 | Sigma=50 | Sigma=70 | Sigma=Blind | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Params(M) | Runtime(s) | FLOPs | PSNR | SSIM | SAM | PSNR | SSIM | SAM | PSNR | SSIM | SAM | PSNR | SSIM | SAM | |
Noisy | 18.59 | 0.110 | .0807 | 14.15 | 0.046 | 0.991 | 11.23 | 0.025 | 1.105 | 17.34 | 0.114 | 0.859 | |||
BM4D | 154 | 38.45 | 0.934 | 0.126 | 35.60 | 0.889 | 0.169 | 33.70 | 0.845 | 0.207 | 37.66 | 0.914 | 0.143 | ||
TDL | 18 | 40.58 | 0.957 | 0.062 | 38.01 | 0.932 | 0.085 | 36.36 | 0.909 | 0.105 | 39.91 | 0.946 | 0.072 | ||
ITSReg | 907 | 41.48 | 0.961 | 0.088 | 38.88 | 0.941 | 0.098 | 36.71 | 0.923 | 0.112 | 40.62 | 0.953 | 0.087 | ||
LLRT | 627 | 41.99 | 0.967 | 0.056 | 38.99 | 0.945 | 0.075 | 37.36 | 0.930 | 0.087 | 40.97 | 0.956 | 0.064 | ||
KBR | 1755 | 41.48 | 0.984 | 0.088 | 39.16 | 0.974 | 0.100 | 36.71 | 0.961 | 0.113 | 40.68 | 0.979 | 0.080 | ||
WLRTR | 1600 | 42.62 | 0.988 | 0.056 | 39.72 | 0.978 | 0.073 | 37.52 | 0.967 | 0.095 | 41.66 | 0.983 | 0.064 | ||
NGmeet | 166 | 42.99 | 0.989 | 0.050 | 40.26 | 0.980 | 0.059 | 38.66 | 0.974 | 0.067 | 42.23 | 0.985 | 0.053 | ||
HSID | 0.40 | 3 | 38.70 | 0.949 | 0.103 | 36.17 | 0.919 | 0.134 | 34.31 | 0.886 | 0.161 | 37.80 | 0.935 | 0.116 | |
QRNN3D | 0.86 | 0.73 | 42.22 | 0.988 | 0.062 | 40.15 | 0.982 | 0.074 | 38.30 | 0.974 | 0.094 | 41.37 | 0.985 | 0.068 | |
TS3C | 0.83 | 0.95 | 42.36 | 0.986 | 0.079 | 40.47 | 0.980 | 0.087 | 39.05 | 0.974 | 0.096 | 41.52 | 0.983 | 0.085 | |
GRUNet | 14.2 | 0.87 | 42.84 | 0.989 | 0.052 | 40.75 | 0.983 | 0.062 | 39.02 | 0.977 | 0.080 | 42.03 | 0.987 | 0.057 |
Complex Denoising on ICVL
non-iid | g+stripe | g+deadline | g+impulse | mixture | ||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Params(M) | Runtime(s) | FLOPs | PSNR | SSIM | SAM | PSNR | SSIM | SAM | PSNR | SSIM | SAM | PSNR | SSIM | SAM | PSNR | SSIM | SAM | |
Noisy | 18.25 | 0.168 | 0.898 | 17.80 | 0.159 | 0.910 | 17.61 | 0.155 | 0.917 | 14.80 | 0.114 | 0.926 | 14.08 | 0.099 | 0.944 | |||
LRMR | 32.80 | 0.719 | 0.185 | 32.62 | 0.717 | 0.187 | 31.83 | 0.709 | 0.227 | 29.70 | 0.623 | 0.311 | 28.68 | 0.608 | 0.353 | |||
LRTV | 33.62 | 0.905 | 0.077 | 33.49 | 0.905 | 0.078 | 32.37 | 0.895 | 0.115 | 31.56 | 0.871 | 0.242 | 30.47 | 0.858 | 0.287 | |||
NMoG | 34.51 | 0.812 | 0.187 | 33.87 | 0.799 | 0.265 | 32.87 | 0.797 | 0.276 | 28.60 | 0.652 | 0.486 | 27.31 | 0.632 | 0.513 | |||
TDTV | 38.14 | 0.944 | 0.075 | 37.67 | 0.940 | 0.081 | 36.15 | 0.930 | 0.099 | 36.67 | 0.935 | 0.094 | 34.77 | 0.919 | 0.113 | |||
HSID | 0.40 | 3 | 38.40 | 0.947 | 0.095 | 37.77 | 0.942 | 0.104 | 37.65 | 0.940 | 0.102 | 35.00 | 0.899 | 0.174 | 34.05 | 0.888 | 0.181 | |
TS3C | 0.83 | 0.95 | 41.12 | 0.986 | 0.069 | 40.66 | 0.985 | 0.077 | 39.38 | 0.982 | 0.100 | 35.92 | 0.951 | 0.205 | 34.36 | 0.945 | 0.230 | |
QRNN3D | 0.86 | 0.73 | 42.79 | 0.978 | 0.052 | 42.35 | 0.976 | 0.055 | 42.23 | 0.976 | 0.056 | 39.23 | 0.945 | 0.109 | 38.25 | 0.938 | 0.107 | |
GRUNet | 14.2 | 0.87 | 42.89 | 0.992 | 0.047 | 42.39 | 0.991 | 0.050 | 42.11 | 0.991 | 0.050 | 40.70 | 0.985 | 0.067 | 38.51 | 0.981 | 0.081 |
If you find this repo helpful, please considering citing us.
@article{LAI2022281,
title = {Deep plug-and-play prior for hyperspectral image restoration},
journal = {Neurocomputing},
volume = {481},
pages = {281-293},
year = {2022},
issn = {0925-2312},
doi = {https://doi.org/10.1016/j.neucom.2022.01.057},
author = {Zeqiang Lai and Kaixuan Wei and Ying Fu},
}