IDN-tensorflow

[Original Caffe version]

Testing

• Install Tensorflow 1.11, Matlab R2017a
• Download Test datasets
• Modify config.py (if you want to test x3 model on Set14, config.TEST.model_path = 'checkpoint_x3/model.ckpt' config.TEST.dataset = 'Set14') and test.py (scale = 3).
• Run testing:

python test.py

Training

• Download Training dataset
• Modify config.py (if you want to train x4 model, config.TRAIN.hr_img_path = '/path/to/DIV2K_train_HR/' config.TRAIN.checkpoint_dir = 'checkpoint_x4/' config.VALID.hr_img_path = '/path/to/DIV2K_valid_HR/' config.VALID.lr_img_path = '/path/to/DIV2K_valid_LR_x4/') and train_SR.py (scale = 4)
• Run training:

python train_SR.py

Note

This TensorFlow version is trained with DIV2K training dataset on RGB channels. Additionally, We modify the upsample layer to subpixel convolution (the original version is transposed convolution).

Results

Test_results

The following PSNR/SSIMs are evaluated on Matlab R2017a and the code can be referred to Evaluate_PSNR_SSIM.m.

Training dataset	Scale	Set5	Set14	B100	Urban100
291	×2	37.83 / 0.9600	33.30 / 0.9148	32.08 / 0.8985	31.27 / 0.9196
DIV2K	×2	37.85 / 0.9598	33.58 / 0.9178	32.11 / 0.8989	31.95 / 0.9266
291	×3	34.11 / 0.9253	29.99 / 0.8354	28.95 / 0.8013	27.42 / 0.8359
DIV2K	×3	34.24 / 0.9260	30.27 / 0.8408	29.03 / 0.8038	27.99 / 0.8489
291	×4	31.82 / 0.8903	28.25 / 0.7730	27.41 / 0.7297	25.41 / 0.7632
DIV2K	×4	31.99 / 0.8928	28.52 / 0.7794	27.52 / 0.7339	25.92 / 0.7801

Model Parameters

Scale	Model size
×2	579,276
×3	587,931
×4	600,048

Citation

If you find IDN useful in your research, please consider citing:

@inproceedings{Hui-IDN-2018,
  title={Fast and Accurate Single Image Super-Resolution via Information Distillation Network},
  author={Hui, Zheng and Wang, Xiumei and Gao, Xinbo},
  booktitle={CVPR},
  pages = {723--731},
  year={2018}
}