Master thesis on natural image noise removal using Convolutional Neural Networks. Works with the Natural Image Noise Dataset to apply to real photographs, using a UNet network architecture by default.

Much lighter version of https://github.com/trougnouf/mthesis-denoise

TODO: C implementation that can be included in darktable, find lighter functional network architecture, reprocess raw data to train model at the beginning of pixel pipeline

See also:

• Dataset: https://commons.wikimedia.org/wiki/Natural_Image_Noise_Dataset
• Paper: https://openaccess.thecvf.com/content_CVPRW_2019/html/NTIRE/Brummer_Natural_Image_Noise_Dataset_CVPRW_2019_paper.html
• Models: https://drive.google.com/drive/folders/1XmY9yO3yhhhdwQ_btYCIpkUBFQ88H-pr?usp=sharing

pytorch, pytorch-opencv, python-configargparse, pyyaml, piqa

eginstall on any operating system: pip3 install --user torch torchvision ConfigArgParse opencv-python pyyaml piqa

eginstall on Arch Linux:

sudo pacman -S python-pytorch-opt-cuda python-opencv
pacaur -Se python-torchvision-cuda  # I highly recommend removing the check() function because the tests take forever
pacaur -S python-pytorch-piqa python-configargparse-git  # configargparse is outdated in the arch repo and results in 'None' (string) values.

python3 denoise_image.py --network <network_architecture> --model_path <model_filepath> --input <input_image_path> [-o output_image_path]

--device can be set to 0 for a CUDA-capable GPU, or -1 for CPU (slow)

the --batch_size can be increased (default: 1) to process multiple patches in parallel (less processing speed, more memory usage).

the output image path might be autogenerated based on the model's filepath.

eg (recommended for most applications): python3 denoise_image.py --network UNet --model_path "../../models/2019-02-18T20:10_run_nn.py_--time_limit_259200_--batch_size_94_--test_reserve_ursulines-red_stefantiek_ursulines-building_MuseeL-Bobo_CourtineDeVillersDebris_MuseeL-Bobo-C500D_--skip_sizecheck_--lr_3e-4" -input <input_image_path> [-o output_image_path]

or if you have a UtNet model (likely even better): python denoise_image.py --network UtNet --model_path <PT_MODEL_PATH> --input <IMAGE_TO_DENOISE>

Note that you should run this on images which have not been sharpened, and apply sharpening then.

The following script tests a model on data it has not seen during training (--test_reserve list), and reports a final ms-ssim, ssim, and mse losses.

python denoise_dir.py --network UtNet --model_path <PT_MODEL_PATH>

(add originally downloaded dataset information with --orig_data, eg ../../datasets/NIND if it is not in its default location)

You can use "--device -1" to perform the test on CPU. Slow but useful if you are already using the GPU for training.

Requirements: python-pytorch, python-configargparse >= 1.3, python-opencv, python-piqa, torchvision, bash, imagemagick, python-opencv, libjpeg[-turbo] [, wget]

eg installation on (Arch Linux):

sudo pacman -S python-pytorch-opt-cuda python-opencv imagemagick libjpeg-turbo wget
pacaur -Se python-torchvision-cuda  # I highly recommend removing the check() function because the tests take forever
pacaur -S python-pytorch-piqa python-configargparse-git  # configargparse is outdated in the arch repo and results in 'None' (string) values.

eg pip installation (note that imagemagick and libjpeg(<-turbo>) are not python packages and need to be installed for your distribution):

pip3 install --user torch torchvision piqa ConfigArgParse 
pip3 install --user opencv-python # recommended to install python-opencv from your distribution's package manager instead
echo 'Don't forget to install imagemagick and libjpeg'

Note that cropping png files is currently extremely slow for non-JPEG images, due to the cropping process opening the whole file every time it makes a new crop. TODO use opencv and open once save many for tiff/png images.

python3 tools/dl_ds_1.py --use_wget         # --use_wget is much less likely to result in half-downloaded files
python3 tools/crop_ds.py --cs 256 --stride 192  # this takes a long time.
python tools/pick_validation_set.py --num_crops 300

Optionally, run python tools/make_dataset_crops_list.py (with the same dataset-related options as above) to generate a list of crops with ms-ssim loss in datasets/-msssim.csv. (Provided for NIND_256_192) This can be useful if you would like to modify the dataset to handle a simpler list and only take images above a given quality threshold.

Cropping can be accelerated greatly by temporarily placing the downloaded dataset (approximately 30.1 GB) on a SSD.

If you would like to train with a different crop size / ucs combination or test set, you must run tools/pick_validation_set.py with the appropriate parameters.

A valid dataset size is determined as follow:

• The minimum size for MS-SSIM is 161 (as determined in common.pt_losses). Account for 25% border with U-Net. MS-SSIM does the equivalent of /16 then conv 11; 176px appears to be an ideal size
• U-Net: conv: -2px, down: /2. -> ((((dim−4)÷2−4)÷2−4)÷2−4)÷2−4 should be whole-> 172, 188, 204, 220. UtNet (not fully tested): bottom: (((176÷2−4)÷2−4)÷2−4)÷2−2 (184 px works)
• JPEG lossless crop: must be divideable by /16 -> crop size of 256 px with 64 px offset; random crop of 220, central loss of

In the end you should have ROOT/datasets/cropped/NIND_256_192// which takes up approximately 46.4 GB. Consider putting that directory on a SSD for faster training.

Requirements: mwclient

bash make_clean-clean_dataset.sh

Unfortunately some of the tools/libraries behind this are not public / oss :( PR for free replacements are welcome.

In the meantime one can either crop /scale a dataset yourself, or even use full-size images.

to use clean data in training, add --clean_data_dpath and --clean_data_ratio (eg: 0.1) argument to nn_train.py

# batch_size 20 is for an 8 GB GPU, adjust for available VRAM
# train a single U-Net generator:
python3 nn_train.py --config configs/train_conf_unet.yaml --batch_size 20 --train_data ../../datasets/train/NIND_256_192

UtNet uses transposed convolutions instead of convolutions in the decoder, resulting in a symmetrical encoder/decoder and no padding necessary; fewer resources and better denoising. This network is recommended.

# batch_size 30 is for an 8 GB GPU, adjust for available VRAM
# train a single UtNet generator:
python3 nn_train.py --config configs/train_conf_utnet_std.yaml --batch_size 30 --train_data ../../datasets/train/NIND_256_192

The beginning of UtNet training is often unstable. If the loss reaches <0.1 then increases back to >0.4 during the first epoch, then the network is stuck producing gargbage; try again. This could probably be solved by starting without the high ISO images, ie by using the PickyDenoisingDatasetFromList for the first epoch, but in the meantime it should initialize correctly within 5-10 attempts (and/or initializing the weights differently).

python3 nn_train.py --d_network Hulf112Disc --batch_size 10

note that run_nn.py contains slightly more options (such as compression and artificial noise) but it only trains one network at a time. nn_train.py can currently train one generator and two discriminators.