jiahong-fu / kxnet

Implementation of 'KXNet: A Model-Driven Deep Neural Network for Blind Super-Resolution' (ECCV2022 Oral)

KXNet: A Model-Driven Deep Neural Network for Blind Super-Resolution (ECCV, 2022 Oral)

Jiahong Fu, Hong Wang, Qi Xie, Qian Zhao, Deyu Meng, and Zongben Xu
​[Paper] [Supplementary Material]
Implementation of 'KXNet: A Model-Driven Deep Neural Network for Blind Super-Resolution' (ECCV2022)

This repo is build on the basis of [DAN] and [IKC]

News

• Add KXNet

Dependenices

• python3
• pytorch >= 1.6
• NVIDIA GPU + CUDA
• Python packages: pip3 install numpy opencv-python lmdb pyyaml

Pretrained models

Pretrained models of KXNet are available as here. Download the models to ./checkpoints

Dataset Preparation

For train We use DIV2K and Flickr2K as our training datasets. Download them and put the images in DIV2K_Flickr2K.

For evaluation, we use four datasets, i.e., Set5, Set14, Urban100, and BSD100. All test datesets can be downloaded from BaiduYun (Password: uag1).

Setting1: We used the Gaussian8 blur kernel set to generate the noise free test data. Run 'codes/scripts/generate_mod_isoblur_LR_sfold.py' to generate LRblur/HR datasets.

Setting2: We separately set the kernel width as $\lambda_1 = 0.8, \lambda_2 = 1.6$ and $\lambda_1 = 2.0, \lambda_2 = 4.0$, and rotate them by $\theta \in { 0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4} }$, respectively. This means we have 8 different blur kernels and use them to generate test data at noise level $\sigma \in { 0, 5, 15}$. Run 'codes/scripts/generate_mod_anisoblur_LR_sfold.py' to generate LRblur/HR datasets.

Train

For single GPU:

cd codes/config/KXNet
python3 train.py -opt=options/setting1/train/train_setting2_x2.yml

For distributed training

cd codes/config/KXNet
python3 -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 train.py -opt=options/setting2/train/train_setting2_x2.yml --launcher pytorch

Test on Synthetic Images

cd codes/config/KXNet
python3 test.py -opt=options/setting2/test/test_setting2_x2.yml

kxnet's Issues

对比方法

感谢您非常优秀的工作。我在重新训练对比方法的时候很不稳定，达不到合理的指标。请问您可以将这三种对比方法的重新训练的checkpoint文件分享给我吗？

文件分享给我吗？

can't find result .

hello,it's amazing work . But when the running is over, I can't find the picture after processing,what should I do to look them?

the Supplementary Material cannot be accessed

[Supplementary Material]

the [Supplementary Material] in your readme cannot be accessed, it shows the 404 Error!

setting1和setting2对于训练过程有什么区别？

您好，在Readme里面看到对于测试集setting1和2不一样，那训练过程有什么不同？为什么要设置setting1和2？

code

Thank you for this great work.

Could you please share the code? thanks

Ask questions in the paper.

Hello！Can you explain how you obtain the gradient of h(X) in Eq. (8) in the paper. Thank you very much.

about Real image test blind model

hello PhD FU, it's an amazing work.How do you test blind models with real images?thank you.Looking forward to your reply.

Paper requirement

Hello, can you provide your full paper? My email: [email protected]. Thx.

The validation datasets

So amazing and meaningful work! I have confused in retraining the KDSR because of the validation datasets. As mentioned in file options, you adpot Urban10 validation datasets during training under both setting 1 and setting 2, is the synthetic 10 images from Urban100 validation datasets? could you please tell the details of which 10 HR and its corresponding LQ images are used for validation. Thanks for your reply. Best wishes for you!

关于测试集生成

您好，请问您setting1生成的测试集是和DAN这篇工作不一样的吗？我用的在DAN上测试集测试您的KXNet，指标结果和您论文里的table1的结果差很多。虽然我看您测试的数据集好像不是lmdb格式，但这应该不是影响测试指标的主要因素吧。