search

mgharbi / demosaicnet_caffe

An implementation of Deep Joint Demosaicking and Denoising - SiGGRAPH Asia 2016
http://groups.csail.mit.edu/graphics/demosaicnet/
Other
110 stars 34 forks source link

readme

This code is outdated, check https://github.com/mgharbi/demosaicnet instead

Deep Joint Demosaicking and Denoising

SiGGRAPH Asia 2016

Michaël Gharbi gharbi@mit.edu Gaurav Chaurasia Sylvain Paris Frédo Durand

Installation and dependencies

This code uses the following external packages:

You can install the python packages via pip: pip install -r requirements.txt

Running

Use the bin/demosaick script, for example:

python bin/demosaick --input data/test_images/000001.png --output output --model pretrained_models/bayer_noise --noise 0.02

For results on the Xtrans mosaick:

python bin/demosaick --input data/test_images/000001.png --output output --model pretrained_models/xtrans --mosaic_type xtrans

Run python bin/demosaick -h for details on the flags you can pass to this script, e.g. --gpu will run the GPU version of the model.

Processing RGB image with ground-truth

When provided with an RGB input, the program will assume it is a ground-truth demosaicked image. It will add noise if requested, mosaick-it, run our algorithm then compute PSNR.

Processing RAW files

When provided with a grayscale image, the program assumes it is a GRBG Bayer mosaic. Use the offset-x and offset-y flag, if you need to shift the mosaick pattern.

python bin/demosaick --input data/test_raw_images/5d_mark_ii_iso800.tiff --output output --model pretrained_models/bayer --offset_x 1

To convert a RAW file to a grayscale Bayer image suitable as input, you can use DCRaw:

dcraw -T -d -W -6 {filename}

This input can then be fed to bin/demosaick

To produce a comparable output from DCRaw's demosaicking algorithm run:

dcraw -T -o 0 -M -W -6 {filename}

Output

When the ground-truth is available, the program outputs a horizontal stack of images with, in order: ground-truth input, noise-corrupted input, corrupted mosaick, denoised/demosaicked result, max-scaled difference map.

If the ground-truth is not available, the program simply outputs a demosaicked image.

Models

We provide three pre-trained Caffe models in the pretrained_models/ directory. bayer has been trained with no noise, and bayer_noise) with noise variances in the range [0, 20] (out of 255). The noise agnostic model will not attempt to perform denoising.

Downloading the dataset.

You can download the full training dataset here. Alternatively, run the download script provided with this code:

cd data
python download_dataset.py

Training a new model

You will first need to generate a new network description by running 

python bin/create_net --output output/new_net

The script is populated with sensible default, but check python bin/create_net -h for details on the parameters you can alter.

Then, generate lmdb training and validation databases from the downloaded datasets:

python bin/convert_to_lmdb --input data/images/train --output data/db_train
python bin/convert_to_lmdb --input data/images/val --output data/db_val

You can now run the training code:

caffe train --solver output/new_net/solver.prototxt --log_dir output/new_net/log

Here's a full example with the dummy database files provided:

python bin/create_net --output output/dummy_net --train_db data/dummy_train_db --test_db data/dummy_val_db
caffe train --solver output/dummy_net/solver.prototxt --log_dir output/dummy_net/log

Running into issues?

Contact Michael Gharbi gharbi@mit.edu

Known issues: