This repository provides a brief guideline of using SAR image colorization network, including preparing dataset, training and test model, and reconstruction.
- More detail information about the network as well as how it works at here
- Contact me at songq15@fudan.edu.cn
- Windows 7
- Python 3.6+
- Tensorflow
- SciPy
- NumPy
- h5py
Two sample data (both smaller than 100M) are used in this repo to show how we organize the data. Note that due to the limited number of training data, the learned model may not be that useful.
- Training: C_SH.mat
- Test: C_NJ.mat
- VGG16 model: download it at here and store it in './data/'.
In this repo, full-pol data is represented by covariance matrix, which has a shape of 3*3*M*N. Use pre_data_tf to extract the full-pol features as well as the input data of the network. Note that for test, only the input data is needed, thus use the function pre_test_data_tf instead.
- To train the model:
python main.py - To test the model:
python main.py --is_train=False --data="./data/data_NJ.mat"
Use the function Recons_from_feature to recover the covariance matrix from predicted full-pol data.
[1] Q. Song, F. Xu, and Y.Q. Jin, "Radar Image Colorization: Converting Single-Polarization to Fully Polarimetric Using Deep Neural Networks," IEEE Access.
[2] G. Larsson, M. Maire, and G. Shakhnarovich, “Learning representations for automatic colorization,” arXiv: 1603.06668, 2016.