QRNN3D

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

🌟 See also the follow up works of QRNN3D:

• DPHSIR - Plug-and-play QRNN3D that solve any HSI restoration task in one model.
• HSDT - State-of-the-art HSI denoising transformer that follows up 3D paradigam of QRNN3D.
• MAN - Improved QRNN3D with significant performance improvement and less parameters.

📉 Performance: QRNN3D < DPHSIR < MAN < HSDT

Highlights

• 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

Prerequisites

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

Quick Start

1. Preparing your training/testing datasets

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

• The train-test split can be found in ICVL_train.txt and ICVL_test_*.txt. (Note we split the 101 testing data into two parts for Gaussian and complex denoising respectively.)

Training dataset

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

Testing dataset

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

2. Testing with pretrained models

• 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

You can also use hsi_eval.py to evaluate quantitative HSI denoising performance.

3. Training from scratch

• 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

Citation

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

@article{wei2020QRNN3D,
  title={3-D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising},
  author={Wei, Kaixuan and Fu, Ying and Huang, Hua},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2020},
  publisher={IEEE}
}

and follow up works

@article{lai2022dphsir,
    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},
}

@inproceedings{lai2023hsdt,
  author = {Lai, Zeqiang and Chenggang, Yan and Fu, Ying},
  title = {Hybrid Spectral Denoising Transformer with Guided Attention},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year = {2023},
}

@article{lai2023mixed,
  title={Mixed Attention Network for Hyperspectral Image Denoising},
  author={Lai, Zeqiang and Fu, Ying},
  journal={arXiv preprint arXiv:2301.11525},
  year={2023}
}

Contact

Please contact me if there is any question (kaixuan.wei at kaust.edu.sa)