Reimplement of Burst Denoising with Kernel Prediction Networks and Multi-Kernel Prediction Networks for Denoising of Image Burst by using PyTorch.
The partial work is following https://github.com/12dmodel/camera_sim.
Write the documents.
- Python3
- PyTorch >= 1.0.0
- Scikit-image
- Numpy
- TensorboardX (needed tensorflow support)
Firstly, you can clone this repo. including train and test codes. Download pretrained model for grayscale images at https://drive.google.com/open?id=1Xnpllr1dinAU7BIN21L3LkEP5AqMNWso, and for color images at https://drive.google.com/file/d/1Il-n7un_u8wWizjQ5ZKQ5hns7S27b0HW/view?usp=sharing.
The repo. supports multiple GPUs to train and validate, and the default setting is multi-GPUs. In other words, the pretrained model is obtained by training on multi-GPUs.
- If you want to restart the train process by yourself, the command you should type is that
CUDA_VISIBLE_DEVICES=x,y train_eval_sym.py --cuda --mGPU -nw 4 --config_file ./kpn_specs/kpn_config.conf --restart
If no option of --restart, the train process could be resumed from when it was broken.
- If you want to evaluate the network by pre-trained model directly, you could use
CUDA_VISIBLE_DEVICES=x,y train_eval_syn.py --cuda --mGPU -nw 4 --eval
If else option -ckpt is choosen, you can select the other models you trained.
- Anything else.
- The code for single image is not released now, I will program it in few weeks.
The following images and more examples can be found at here.
 |
 |
 |
| Ground Truth | Noisy | Denoised |
 |
 |
 |
| Ground Truth | Noisy | Denoised |
The following images and more examples can be found at here.
 |
 |
 |
| Ground Truth | Noisy (PSNR: 19.34dB, SSIM: 0.595) | Denoised (PSNR: 29.69dB, SSIM: 0.937) |
 |
 |
 |
| Ground Truth | Noisy (PSNR: 18.70dB, SSIM: 0.308) | Denoised (PSNR: 34.02dB, SSIM: 0.954) |