MAFNet

The implementation of IEEE JSTARS 2023 paper "Multi-scale Adaptive Fusion Network for Hyperspectral Image Denoising"

Results

• Comparison of Quantitative Indicators

• Comparison of Visual Quality

Configuration

• Python =3.6.8, PyTorch = 1.9.0
• Requirements: listed in requirements.txt
• Platforms: Ubuntu 16.04.7 LTS x86_64, cuda-10.2

Quick Start

1. Prepare training/testing datasets

Training dataset

• Download ICVL hyperspectral image database

• Put the downloaded .mat file under the data/Mat_icvl, and it will be cropped with the specified step size to obtain training and validation dataset under the dataset_p_icvl.
• The operation method of other data sets is the same as above

Testing dataset

• Put the test data under the data/test path, read and modify the test.py file for model testing.

2. Testing with pretrained models

• The trained model will be stored in the model path.

• [Gaussian noise removal]:
  python3 test.py -pm './model/MAFNet_icvl_guass.pkl' -m 'MAFNet' -b 'no'-n 50

• [Complex noise removal]:
  python3 test.py -pm './model/MAFNet_icvl_case.pkl' -m 'MAFNet' -b 'case5'

3. Training with incremental learning policy

• Training a blind Gaussian model firstly by
  python3 main.py -d './data/dataset_p_icvl/' -mp './data/Mat_icvl/' -m 'MAFNet' -e 150 -b 'guass' -c 0 -dn 'icvl'

• Using the pretrained Gaussian model as initialization to train a complex model:
  python3 main.py -d './data/dataset_p_icvl/' -mp './data/Mat_icvl/' -m 'MAFNet' -e 150 -b 'case5' -pm './model/MAFNet_icvl_guass.pkl' -c 0 -dn 'icvl'

Citation

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

@ARTICLE{mafnet23jstars,
  author={Pan, Haodong and Gao, Feng and Dong, Junyu and Du, Qian},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={Multi-Scale Adaptive Fusion Network for Hyperspectral Image Denoising}, 
  year={2023},
  pages={1-16},
  doi={10.1109/JSTARS.2023.3257051}}

Contact

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