The implementation of TNNLS 2020 paper "3D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising"

• Our network outperforms all leading-edge methods (2019) on ICVL dataset in both Gaussian and complex noise cases, as shown below:

• We demonstrated our network pretrained on 31-bands natural HSI database (ICVL) can be utilized to recover remotely-sensed HSI (> 100 bands) corrupted by real-world non-Gaussian noise due to terrible atmosphere and water absorptions

• Python >=3.5, PyTorch >= 0.4.1
• Requirements: opencv-python, tensorboardX, caffe
• Platforms: Ubuntu 16.04, cuda-8.0

Download ICVL hyperspectral image database from here (we only need .mat version)

Note cafe (via conda install) and lmdb are required to execute the following instructions.

• Read the function create_icvl64_31 in utility/lmdb_data.py and follow the instruction comment to define your data/dataset address.

• Create ICVL training dataset by python utility/lmdb_data.py

Note matlab is required to execute the following instructions.

• Read the matlab code of matlab/generate_dataset* to understand how we generate noisy HSIs.

• Read and modify the matlab code of matlab/HSIData.m to generate your own testing dataset

• Download our pretrained models from OneDrive and move them to checkpoints/qrnn3d/gauss/ and checkpoints/qrnn3d/complex/ respectively.

• [Blind Gaussian noise removal]:
  python hsi_test.py -a qrnn3d -p gauss -r -rp checkpoints/qrnn3d/gauss/model_epoch_50_118454.pth

• [Mixture noise removal]:
  python hsi_test.py -a qrnn3d -p complex -r -rp checkpoints/qrnn3d/complex/model_epoch_100_159904.pth

• Training a blind Gaussian model firstly by
  python hsi_denoising_gauss.py -a qrnn3d -p gauss --dataroot (your own dataroot)

• Using the pretrained Gaussian model as initialization to train a complex model:
  python hsi_denoising_complex.py -a qrnn3d -p complex --dataroot (your own dataroot) -r -rp checkpoints/qrnn3d/gauss/model_epoch_50_118454.pth --no-ropt

If you find this work useful for your research, please cite:

@article{wei2020QRNN3D,
  title={3D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising},
  author={Wei, Kaixuan and Fu, Ying and Huang, Hua},
  journal={arXiv preprint arXiv:2003.04547},
  year={2020}
}

Please contact me if there is any question (Kaixuan Wei [email protected])