Burst Denoising of Dark Images

This repo is an unofficial Pytorch implementation of Burst Denoising of Dark Images (Link) by Ahmet Serdar Karadeniz, Erkut Erdem, Aykut Erdem

Requirement

* pytorch
* rawpy
* tqdm
* torchsummary
* numpy
* skimage
* tensorboard
* opencv-python 

Dataset (SID)

• SID dataset contains two type of RAW images, one from Sony camera and another from Fuji camera, this repo uses Sony part of SID dataset for training the models.
• Link: Sony(25GB), Fuji(52GB) [2]
• Download and extract to folder dataset/
• Remember to modify input channel of two models if you want to train model for Fuji subset.

Link to pretrained model:

Link

Train

1. Single frame model

1.1. Train Coarse model

python train_coarse.py train -e <experiment_name> 

1.2. Train Fine model

• If we understand correctly the author's idea, fine model reuses trained weights from coarse model, so please adjust the code of train_fine.py to correct location of pretrained coarse model (gotten from step 1.1)

• In order to save time for preprocessing while training, this repo uses preprocessed ground truth images Sony GT. Download and extract it to folder dataset/gt/

python train_fine.py train -e <experiment_name> -c <coarse_checkpoint>

2. Set-based model

• currently not supported

Test

Testing trained model on dataset/Sony_test_list.txt

python test.py -c <coarse_checkpoint_dir> -f <fine_checkpoint_dir> -s <saved_folder>

Sample inference

check out notebook file: inference_sample.ipynb

References:

• [1] Karadeniz, A.S., Erdem, E. and Erdem, A., 2020. Burst Denoising of Dark Images. arXiv preprint arXiv:2003.07823. https://arxiv.org/abs/2003.07823v1
• [2] https://github.com/cchen156/Learning-to-See-in-the-Dark
• [3] https://github.com/cydonia999/Learning_to_See_in_the_Dark_PyTorch