Unpaired Image Denoising

Code for the paper: Unpaired Image Denoising (accepted for ICIP 2020).

Instructions to reproduce

Module dependencies are listed in requirements.txt.

First download the MS COCO dataset and split it into 3 folders: clean, noisy and test.

Stage 1

In this stage, the images in clean are used to train a flow-based model.

cd src/glow
python train.py --clean_path=<Path to `clean` split of COCO dataset>

Stage 2

In this stage, a resnet is trained with inputs from the noisy folder with the flow-based model trained above as prior.

cd src/
python train.py --datsaset=<Path to full COCO dataset> --saved_flow_model=<Path to ckpt file of trained flow based model>

Testing is also done along with training (see src/train.py).