Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, Timo Aila

Abstract:

We apply basic statistical reasoning to signal reconstruction by machine learning -- learning to map corrupted observations to clean signals -- with a simple and powerful conclusion: it is possible to learn to restore images by only looking at corrupted examples, at performance at and sometimes exceeding training using clean data, without explicit image priors or likelihood models of the corruption. In practice, we show that a single model learns photographic noise removal, denoising synthetic Monte Carlo images, and reconstruction of undersampled MRI scans -- all corrupted by different processes -- based on noisy data only.

For business inquiries, please visit our website and submit the form: NVIDIA Research Licensing

• Paper (arXiv)

All the material, including source code, is made freely available for non-commercial use under the Creative Commons CC BY-NC 4.0 license. Feel free to use any of the material in your own work, as long as you give us appropriate credit by mentioning the title and author list of our paper.

The below sections detail how to get set up for training the Noise2Noise network using the ImageNet validation dataset. Noise2Noise MRI denoising instructions are at the end of this document.

This code is tested with Python 3.6. We're using Anaconda 5.2 to manage the Python environment. Here's how to create a clean environment and install library dependencies:

conda create -n n2n python=3.6
conda activate n2n
conda install tensorflow-gpu
python -m pip install --upgrade pip
pip install -r requirements.txt

This will install TensorFlow and other dependencies used in this project.

This section explains how to prepare a dataset into a TFRecords file for use in training the Noise2Noise denoising network. The image denoising results presented in the Noise2Noise paper have been obtained using a network trained with the ImageNet validation set.

Training dataset for ImageNet: To generate the ImageNet validation set tfrecords file, run:

# This should run through roughly 50K images and output a file called `datasets/imagenet_val_raw.tfrecords`.
python dataset_tool_tf.py --input-dir "<path_to_imagenet>/ILSVRC2012_img_val" --out=datasets/imagenet_val_raw.tfrecords

A successful run of dataset_tool_tf.py should print the following upon completion:

<...long omitted...>
49997 ./ImageNet/ILSVRC2012_img_val/ILSVRC2012_val_00002873.JPEG
49998 ./ImageNet/ILSVRC2012_img_val/ILSVRC2012_val_00031550.JPEG
49999 ./ImageNet/ILSVRC2012_img_val/ILSVRC2012_val_00009765.JPEG
Dataset statistics:
  Formats:
    RGB: 49100 images
    L: 899 images
    CMYK: 1 images
  width,height buckets:
    >= 256x256: 48627 images
    < 256x256: 1373 images

Training dataset for BSD300: If you don't have access to the ImageNet training data, you can experiment with this code using the BSD300 dataset -- just note that the results will be slightly worse than with a network trained on ImageNet.

Download the BSD300 dataset images to some folder. The below example assumes the dataset has been uncompressed into the ./datasets directory.

Convert the images in BSD300 training set into a tfrecords file:

python dataset_tool_tf.py --input-dir datasets/BSDS300-images/BSDS300/images/train --out=datasets/bsd300.tfrecords

Kodak validation set. Training tests validation loss against the Kodak Lossless True Color Image Suite dataset. Here's how to prepare this dataset for use during training:

# Download the kodak validation set from http://r0k.us/graphics/kodak/
python download_kodak.py --output-dir=datasets/kodak

To train the noise2noise autoencoder on ImageNet:

# try python config.py train --help for available options
python config.py --desc='-test' train --train-tfrecords=datasets/imagenet_val_raw.tfrecords --long-train=true --noise=gaussian

You can inspect the training process using TensorBoard:

cd results
tensorboard --logdir .

By default, this invocation will train a Gaussian denoising network using the ImageNet validation set. This takes roughly 7.5 hours on an NVIDIA Titan V GPU.

Upon completion, the training process produces a file called network_final.pickle under the results/* directory.

Once you've trained a network, you can run a validation dataset through the network:

Suppose your training run results were stored under results/00001-autoencoder-1gpu-L-n2n. Here's how to run a set of images through this network:

python config.py validate --dataset-dir=datasets/kodak --network-snapshot=results/00001-autoencoder-1gpu-L-n2n/network_final.pickle

You can find pre-trained networks for Poisson and Gaussian noise removal here: https://drive.google.com/drive/folders/1-84ORv4wB8W3M6WngFTtccuW7SlPku0V

Here's a summary of training options to reproduce results from the Noise2Noise paper:

To validate against a trained network, use the following options:

Note: When running a validation set through the network, you should match the augmentation noise (e.g., Gaussian or Poisson) with the type of noise that was used to train the network.

Use the dataset_tool_mri.py script to generate training and validation datasets for the N2N MRI case.

Step #1: Download the IXI-T1 dataset from: https://brain-development.org/ixi-dataset/. Unpack to some location.

Step #2: Convert the IXI-T1 dataset into a set of PNG files:

# Assumes you have downloaded and untarred IXI-T1 under ~/Downloads/IXI-T1.

python dataset_tool_mri.py genpng --ixi-dir=~/Downloads/IXI-T1 --outdir=datasets/ixi-png

Step #3: Convert a subset of the IXI-T1 dataset into training and validation sets:

python dataset_tool_mri.py genpkl --png-dir=datasets/ixi-png --pkl-dir=datasets

python config_mri.py

A successful invocation should output the following:

dnnlib: Running train_mri.train() on localhost...
Loading training set.
Loading dataset from datasets\ixi_train.pkl
<...long log omitted...>
Epoch 297/300: time=107.981, train_loss=0.0126064, test_db_clamped=31.72174, lr=0.000002
Epoch 298/300: time=107.477, train_loss=0.0125972, test_db_clamped=31.73622, lr=0.000001
Epoch 299/300: time=106.927, train_loss=0.0126012, test_db_clamped=31.74232, lr=0.000001
Saving final network weights.
Resetting random seed and saving a bunch of example images.
dnnlib: Finished train_mri.train() in 8h 59m 19s.

The expected average PSNR on the validation set (named test_db_clamped in code) is roughly 31.74 dB.

Noise-to-noise training is enabled by default for the MRI case. To use noise-to-clean training, edit config_mri.py and change corrupt_targets=True to corrupt_targets=False.

Training for 300 epochs takes roughly 9 hours on an NVIDIA Titan V GPU.