Paper | Project | pretrained models
- Python >= 3.8 (Recommend to use Anaconda or Miniconda)
- PyTorch >= 1.10
-
Clone repo
git clone git@github.com:Zj-BinXia/BBCU.git
-
If you want to train or test BBCU for super-resolution
cd BBCU-SR -
If you want to train or test BBCU for denoising and deblocking
cd BBCU-denoiseAndblocking
It is notable that our amplification factor k for residual alignment is used to balance the value range gap of full-precision residual branch and binarized Conv branch as input image range is 0-1. The best k∗ is related to the number of feature channels n, which empirically fits k∗ = 130n/64. You can adjust it according to your network setting.
More details please see the README in folder of BBCU-SR and BBCU-denoiseAndblocking
@article{xia2022basic,
title={Basic Binary Convolution Unit for Binarized Image Restoration Network},
author={Xia, Bin and Zhang, Yulun and Wang, Yitong and Tian, Yapeng and Yang, Wenming and Timofte, Radu and Van Gool, Luc},
journal={ICLR},
year={2023}
}
If you have any question, please email zjbinxia@gmail.com.