Suppression of Independent and Correlated Noise with Similarity-based Unsupervised Deep Learning

Under review. (In Updating)

Introduction

This is a general similarity-based unsupervised deep denoising approach to suppress not only independent but also correlated image noise. Theoretical analysis proves the equivalent effectiveness of this unsupervised approach to the supervised counterpart. This project shows several application cases, including denoising natural, microscopic, low-dose CT and photon-counting micro-CT images. This general approach can be applied to many other fields by constructing the similar training samples based on the domain-specific prior.

Installation

Assuming Anaconda with python 3.6, the required packages for this project can be installed as:

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch  # install pytorch
conda install -c pytorch faiss-gpu  # install faiss-gpu
conda install -c conda-forge python-lmdb tqdm imageio addict opencv matplotlib pydicom scipy
pip install mat73

Then, clone this repo

git clone https://github.com/niuchuangnn/noise2sim.git
cd noise2sim

Applications

Natural Images

Download BSD68 test dataset at here, and put them under the folder ./datasets/

Prepare dataset:

python ./tools/prepare_bsd400_lmdb.py

Run on 1 GPU:

python ./tools/train_dist.py --config-file ./configs/bsd400_unet2_ps3_ns8_gpu1.py

Run on 8 GPUs:

python ./tools/train_dist.py --config-file ./configs/bsd400_unet2_ps3_ns8_gpu8.py

The results in the paper were obtained using 8 GPUs, you can obtain similar results with 1 GPU.

Low-dose CT Images

The low-dose CT dataset can be obtained at Low Dose CT Grand Challenge.

Arrange the Mayo data like:

├── datasets   
    ├── Mayo                   
        ├── L067                    
        ├── L096
        ...

Run on 4 GPUs:

Prepare train and test data files:

python tools/prepare_mayo.py --patient-folder L067 L096 L109 L143 L192 L286 L291 L310 --output-file ./datasets/Mayo/mayo_train.txt
python tools/prepare_mayo.py --patient-folder L506 L333 --output-file ./datasets/Mayo/mayo_test.txt

Run on 4 GPUs:

python ./tools/train_dist.py --config-file ./configs/mayo_unet2.py

Spectral CT Images

The photon-counting spectral micro-CT data can be obtained here, and put it under ./datasets/.

Run on 4 GPUs:

python ./tools/train_dist.py --config-file ./configs/sct_unet2.py

Microscopy Images

The microscopy image can be obtained here.

Citation

@inproceedings{noise2sim2021,
  title={Noise2Sim – Similarity-based Self-Learning for Image Denoising},
  author={Niu, Chuang and Wang, Ge},
  booktitle={arXiv:2011.03384},
  year={2020}
}