This is the implementation of our noise classifiers.
For more information, please check our paper in Section 6.
Our experiments were done over sRGB patches from the SIDD (small version). In our experiments, the SIDD images were first split into ISO values, then train/val/test, following by cropping into 256x256 patches. For using the scripts, the code must be changed to account for the directories of the patches.
For installing the requirements, just run:
pip install -r requirements.txt
We have the Keras and Tensorflow versions fixed to the ones we've used. If you use different versions, you will probably need to adjust the code.
File noise_artXreal.py trains and evaluates a network for discriminating artificial noise from natural noise.
This script compares existing models (Gaussian, Poisson, GaussianPoissonian, and Gaussian through CFA+demosaicing) against natural noise.
We can clearly see how easy it is for discriminating artificial from natural noise.
File noise_artXrealXGAN_SIDDXnoiseflow.py trains a classifier for discriminating the above mentioned methods, besides NoiseFlow and ours.
If you wish to download our trained models (for this experiment), and extract them into trained_models. You can do that by running the following bash:
bash ./download_trained_models.shBy running the test_artXrealXGAN_SIDDXnoise_flow, you can test the trained models in sample images (check sample_imgs folder):
python test_artXrealXGAN_SIDDXnoise_flow.py
This will generate a confusion matrix and TSNE projections for the ISO 800.