Home

Awesome

Hyperspectral Image Spatial Super-Resolution via 3D Full Convolutional Neural Network

By Shaohui Mei, Xin Yuan, Jingyu Ji , Yifan Zhang, Shuai Wan, Qian Du


Introduction

The 3D-FRCNN is an unified framework for hyperspectral image super-resolutiom(SR) with a single network. You can use the code to train/evaluate a network for hsi super-resolution(SR). For more details, please refer to our paper.

Proposed Framework

Some SR results


Citing our work

@Article{rs9111139,
AUTHOR = {Mei, Shaohui and Yuan, Xin and Ji, Jingyu and Zhang, Yifan and Wan, Shuai and Du, Qian},
TITLE = {Hyperspectral Image Spatial Super-Resolution via 3D Full Convolutional Neural Network},
JOURNAL = {Remote Sensing},
VOLUME = {9},
YEAR = {2017},
NUMBER = {11},
ARTICLE NUMBER = {1139},
URL = {http://www.mdpi.com/2072-4292/9/11/1139},
ISSN = {2072-4292},
DOI = {10.3390/rs9111139}
}

Installation

Install Keras
  1. Please follow Tensorflow instruction to install all necessary packages and build it.
  2. Please follow Keras instruction

Traineval

datasets
  1. download the datasets from here and save in 'data' folder in data_process' folder
  2. normalized and get mirrore of original data
cd ./data_process
python expand.py
  1. turn the data into small pieces open your matlab and run gen_train_all_bands.m in 'data_process' folder

  2. prepare the train data for keras

cd ./data_process
python get_to_train.py
training
python train_network.py

Test

We release one pretrained models: model_pa.h5 in "model" folder for Pavia dataset. Do testing

python predict.py

will show both reconstructed images and PSNR/SSIM/SAM.