Awesome

Fast Nearest Convolution for Real-Time Image Super-Resolution, AIM & ECCV Workshops 2022, [Paper]

Update

[2022.11.14] A more flexible implementation of the nearest convolution initialization is provided in this issue.

[2022.11.12] We provide a simple implementation of NCNet model using PyTorch in this torch_code directory.

[2022.08.25] We have uploaded the pretrained model in Releases as V1.0

Dependencies

• OS: Ubuntu 18.04
• Python: Python 3.7
• Tensorflow 2.9.1
• nvidia :
  • cuda: 11.2
  • cudnn: 8.1.0
• Other reference requirements

Performance of our Nearest Convolution

Model Training

python main.py

Then the trained keras model will be saved in ckpt/basenet/model folder.

Model Validation

python eval.py

Then the results of original keras model will be saved in original_output folder and you can calculate the validation PSNR by run python calculate_PSNR.py

Convert to TFLite

python generate_tflite.py

Then the converted tflite model will be saved in TFLite folder.

TFLite Model Validation

python test_tflite.py

Then the results of TFLite model will be saved in results folder.

Other Details

• The input image range is [0, 255].
• Number of parameters: 52,279 (53K)
• Average PSNR on DIV2K validation data: 30.27 dB
• Training data: DIV2K.

Citations

If this repo helps your research or work, please consider citing our work. The following is a BibTeX reference.

@inproceedings{luo2023fast,
  title={Fast nearest convolution for real-time efficient image super-resolution},
  author={Luo, Ziwei and Li, Youwei and Yu, Lei and Wu, Qi and Wen, Zhihong and Fan, Haoqiang and Liu, Shuaicheng},
  booktitle={Computer Vision--ECCV 2022 Workshops: Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part II},
  pages={561--572},
  year={2023},
  organization={Springer}
}

Contact

email: [ziwei.ro@gmail.com]