TSCN

Code of "Two-Stage Convolutional Network for Image Super-Resolution" (ICPR 2018)

The schematics of the proposed Two-Stage Convolutional Network.

Architecture of the multipath information fusion module.

Speed and accuracy trade-off. (x3 on Set5)

Run test

Install Caffe, Matlab R2017a
Run testing:

$ cd ./test
$ matlab -nodisplay
>> test_TSCN

The training dataset is 291 images.

The results are stored in "results" folder, with both reconstructed images and PSNR/SSIM/IFCs.

Results

_Method	_Scale	_Set5	_Set14	_B100	_Urban100
_DRRN	_×2	_37.74/0.9591	_33.23/0.9136	_32.05/0.8973	_31.23/0.9188
_TSCN	_×2	_37.88/0.9602	_33.28/0.9147	_32.09/0.8985	_31.29/0.9198
_DRRN	_×3	_34.03/0.9244	_29.96/0.8349	_28.95/0.8004	_27.53/0.8378
_TSCN	_×3	_34.18/0.9256	_29.99/0.8351	_28.95/0.8012	_27.46/0.8362
_DRRN	_×4	_31.68/0.8888	_28.21/0.7721	_27.38/0.7284	_25.44/0.7638
_TSCN	_×4	_31.82/0.8907	_28.28/0.7734	_27.42/0.7301	_25.44/0.7644

Train

step 1: Compile Caffe with train/include/caffe/layers/l1_loss_layer.hpp, train/src/caffe/layers/l1_loss_layer.cpp and train/src/caffe/layers/l1_loss_layer.cu
step 2: Run data_aug.m to get the augmented 291 dataset
step 3: Run generate_train_TSCN.m to convert training images to hdf5 file
step 4: Run generate_test_TSCN.m to convert testing images to hdf5 file for valid model during the training phase
step 5: Run train.sh to train ×2 model (Manually create directory caffemodel_x2)

Note: You can train the stage-one model by run `train_stage_one.sh`

Citation

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

@inproceedings{Hui-TSCN-2018,
  title={Two-Stage Convolutional Network for Image Super-Resolution},
  author={Hui, Zheng and Wang, Xiumei and Gao, Xinbo},
  booktitle={ICPR},
  pages={2670--2675},
  year={2018}
}

Zheng222/TSCN