https://github.com/czbiohub-sf/noise2self
A framework for blind denoising with self-supervision.
Science Score: 10.0%
This score indicates how likely this project is to be science-related based on various indicators:
-
○CITATION.cff file
-
○codemeta.json file
-
○.zenodo.json file
-
○DOI references
-
✓Academic publication links
Links to: arxiv.org -
○Academic email domains
-
○Institutional organization owner
-
○JOSS paper metadata
-
○Scientific vocabulary similarity
Low similarity (11.1%) to scientific vocabulary
Repository
A framework for blind denoising with self-supervision.
Basic Info
- Host: GitHub
- Owner: czbiohub-sf
- License: mit
- Language: Jupyter Notebook
- Default Branch: master
- Size: 10.4 MB
Statistics
- Stars: 343
- Watchers: 17
- Forks: 74
- Open Issues: 5
- Releases: 0
Metadata Files
README.md
Noise2Self: Blind Denoising by Self-Supervision
This repo demonstrates a framework for blind denoising high-dimensional measurements, as described in the paper. It can be used to calibrate classical image denoisers and train deep neural nets; the same principle works on matrices of single-cell gene expression.
The result of training a U-Net to denoise a stack of noisy Chinese characters. Note that the only input is the noisy data; no ground truth is necessary.
Images
The notebook Intro to Calibration shows how to calibrate any traditional image denoising model, such as median filtering, wavelet thresholding, or non-local means. We use the excellent scikit-image implementations of these methods, and have submitted a PR to incorporate self-supervised calibration directly into the package. (Comments welcome on the PR!)
The notebook Intro to Neural Nets shows how to train a denoising neural net using a self-supervised loss, on the simple example of MNIST digits. The notebook runs in less than a minute, on CPU, on a MacBook Pro. We implement this in pytorch.
The notebook Single Shot Denoising demonstrates that there is enough information in a single 512x512 noisy image for a deep neural net to learn to denoise it, with performance better than classical blind image denoisers.
Because the self-supervised loss is much easier to implement than the data loading, GPU management, logging, and architecture design required for handling any particular dataset, we recommend that you take any existing pipeline for your data and simply modify the training loop.
Traditional Supervised Learning
for i, batch in enumerate(data_loader):
noisy_images, clean_images = batch
output = model(noisy_images)
loss = loss_function(output, clean_images)
Self-Supervised Learning
from mask import Masker
masker = Masker()
for i, batch in enumerate(data_loader):
noisy_images, _ = batch
input, mask = masker.mask(noisy_images, i)
output = model(input)
loss = loss_function(output*mask, noisy_images*mask)
Dependencies are in the environment.yml file.
The remaining notebooks generate figures from the paper.
Owner
- Name: Chan Zuckerberg Biohub San Francisco
- Login: czbiohub-sf
- Kind: organization
- Location: San Francisco
- Website: https://www.czbiohub.org/sf/
- Repositories: 1
- Profile: https://github.com/czbiohub-sf
GitHub Events
Total
- Watch event: 21
- Fork event: 6
Last Year
- Watch event: 21
- Fork event: 6
Dependencies
- ipython
- jupyter
- matplotlib
- numpy
- pillow
- pytorch
- scikit-image
- scipy
- six
- torchvision