https://github.com/johnpertoft/noise2noise

Noise2Noise implementation
https://github.com/johnpertoft/noise2noise

noise2noise tensorflow

Last synced: about 1 month ago
JSON representation

Noise2Noise implementation

• Host: GitHub
• URL: https://github.com/johnpertoft/noise2noise
• Owner: johnPertoft
• Created: 2018-09-06T13:59:42.000Z (almost 8 years ago)
• Default Branch: master
• Last Pushed: 2020-08-26T14:00:42.000Z (almost 6 years ago)
• Last Synced: 2025-03-22T08:48:45.255Z (over 1 year ago)
• Topics: noise2noise, tensorflow
• Language: Python
• Size: 4.33 MB
• Stars: 3
• Watchers: 2
• Forks: 1
• Open Issues: 2
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Noise2Noise

Tensorflow [Noise2Noise](https://arxiv.org/abs/1803.04189) implementation.

Noise2Noise is a machine learning algorithm that can learn signal reconstruction from only

noisy examples, i.e. both inputs and targets are noisy realisations of the same image.

## Prerequisites

* Tfrecord files with jpeg encoded images under key `image/encoded` for training and evaluation.

## Docker

(requires [nvidia-docker](https://github.com/NVIDIA/nvidia-docker))

Build docker image

```bash

$ docker build -t n2n .

```

Run a command inside docker container

```bash

$ ./scripts/run-in-docker 

```

Mount extra volumes for input or output reasons. Current directory is already shared.

```bash

$ VOLUMES="/vol1:/vol1 /vol2:/vol2" ./scripts/run-in-docker 

```

## Help

```bash

$ python -m n2n.train --helpfull

```

## Results

Images from left to right are input image, denoised image, and ground truth noise free image.

### Additive gaussian noise

```bash

$ python -m n2n.train  --noise additive_gaussian --loss l2

```

![additive-gaussian-noise](results/additive-gaussian-2.png)

![additive-gaussian-noise](results/additive-gaussian-gt-loss.png) ![additive-gaussian-noise](results/additive-gaussian-loss.png)

#### Experiment with additional adversarial loss

TODO: Compare on similar images.

```bash

$ python -m n2n.train  --noise additive_gaussian --loss l2 --adv_loss lsgan

```

![additive-gaussian-noise-adv](results/additive-gaussian-adv-loss.png)

### Text overlay noise

```bash

$ python -m n2n.train  --noise text --loss l1

```

![text-noise](results/text.png)

![text-noise](results/text-gt-loss.png) ![text-noise](results/text-loss.png)

### Impulse noise

```bash

$ python -m n2n.train  --noise impulse --loss l0

```

### Bernoulli noise

TODO

### Poisson noise

TODO

## TODO

* Raytracing/raycasting noise?