WMCNN-Pytorch

The Pytorch reproduction of WMCNN [Aerial Image Super Resolution via Wavelet Multiscale Convolutional Neural Networks] If you use this code, please cite the paper.

The PSNR value on RSSCN7 dataset is compared in the following table.

Methods	Upscaling factor	Grass	Field	Industry	River Lake	Forest	Resident	Parking	Average
WMCNN_paper	2	38.82	37.30	28.35	32.41	29.68	28.49	29.10	32.02
WMCNN_pytorch	2	38.98	37.38	28.28	32.31	29.71	28.33	30.00	32.14

Usage

Generate data

First, you need to download the RSSCN7 dataset in this site and put it in the directory "data/rsscn7". Then you can either use the following two methods to generate the hdf5 dataset. (*Note: using other datasets is also possible.)

Matlab

Use the code "generate_train.m" provided in folder 'matlab_generate_data' to generate hdf5 dataset.

Python

If the matlab is not available, you can use python code "data_generator.py" to generate hdf5 dataset.

Training

First, update the dataset path in config file 'configs/wmcnn.json'

 "train_data_loader": {
    "data_path": "./data/rsscn7/",
    "train_path": "full_wdata.h5",
    ...
    }

Use the following code for training,

python main_train.py -c configs/wmcnn.json

After training, the config file will be copied to 'experiments/wmcnn/wmcnn.json' You can check the log during training in 'experiments/wmcnn/log' with tensorboard.

Testing

First, update the testing dataset path in config file 'experiments/wmcnn/wmcnn.json'

  "test_data_loader": {
    "data_path": "./data/rsscn7/",
    "test_path": "Samples/",
    "upscale": 2
  }

Use the following code for testing,

python main_test.py -c experiments/wmcnn/wmcnn.json