An implementation of RCAN described in the paper using tensorflow. Image Super-Resolution Using Very Deep Residual Channel Attention Networks

Published in ECCV 2018, written by Y. Zhang, K. Li, L. Wang, B. Zhong, and Y. Fu

• Python 3.6.5
• Tensorflow 1.13.1
• Pillow 6.0.0
• numpy 1.15.0
• scikit-image 0.15.0

• DIV2K
• Benchmarks

• RCAN_X2
• RCAN_X3
• RCAN_X4

python main.py --train_GT_path ./GT_path --train_LR_path ./LR_path --test_GT_path ./test_GT_path --test_LR_path ./test_LR_path --test_with_train True --scale 2(or 3, 4, ...) --log_freq 1000

• LR image and HR image pair should have same index when they are sorted by name respectively.
• You can refer to the script file (run.sh) in my repository

1. Download pre-trained model.

• RCAN_X2
• RCAN_X3
• RCAN_X4

2. Unzip the pre-trained model file

tar -cvf model.tar

3. Test using benchmarks

python main.py --mode test --pre_trained_model ./model/RCAN_X2(or 3, 4) --test_LR_path ./benchmark_LR_path --test_GT_path ./benchmark_GT_path --scale 2(or 3, 4) --self_ensemble False

If you want to use self_ensemble, --self_ensemble option to True

• You can refer to the script file (run.sh) in my repository

1. Download pre-trained model.

• RCAN_X2
• RCAN_X3
• RCAN_X4

2. Unzip the pre-trained model file

tar -cvf model.tar

3. Inference your own images

python main.py --mode test_only --pre_trained_model ./model/RCAN_X2(or 3, 4) --test_LR_path ./your_own_images --scale 2(or 3, 4) --chop_forward False

If your images are too large, OOM error can occur. In that case, --chop_forward option to True

Qualitative results are will be updated soon!

If you have any questions or comments on my codes, please email to me. [email protected]

[1] https://github.com/yulunzhang/RCAN