Denoisenet

arXiv:1701.01698 by Remez et el.

Result

Usage

Install Requirements

pip install -r requirements.txt

Evaluate with pretrained model:

python3 denoise_eval.py ./data/n000.tif ./data/c000.tif ./Model/denoisenet/denoise_model_I404000.ckpt-404000.meta ./Model/denoisenet/

Where ./Model/denoisenet/denoise_model_I404000.ckpt-404000.meta is tensorflow graph definition, ./Model/denoisenet/ is where checkpoint file locates.

See more with python3 denoise_eval -h

Experiment

Model is trained on house15K dataset for 24 hours on an AWS g2.8xlarge instance. Training is optimized for multi-gpu and achieved 0.8s/it at minibatch size of 16.

Train with your own dataset

In order to train with your machine, please change device layout in the code.

Loading your dataset

2 ways of providing training data is provided:

• In two separate folders, e.g. Clean/ and Noisy/, clean and noisy images are lined up in the same order. For example:

  Clean/

  Clean/001.tif

  Clean/002.tif

  Clean/003.tif

  ...

  Noisy/

  Noisy/001.tif

  Noisy/002.tif

  Noisy/003.tif

  ...

Where Clean/001.tif and Noisy/001.tif are groundtruth and corrupted image of the same training example. Run this line to train.

python3 denoise_train.py -cpr /path/to/Clean/ -npr /path/to/Noisy/

• Wrap your training data in a tfrecords file. This is useful when you have an extra large dataset, and cannot fit entire dataset into memory. With the training data lined up as mentioned in the first method, run this line to start converting:

python3 denoise_input.py /path/to/CleanAndNoisy/ --ycbcr

--ycbcr converts your images into ycbcr format.
Then run this line to train:

python3 denoise_train.py -df /path/to/dataset.tfrecords