This repository provides the codes for training and testing the Noise Flow model used for image noise modeling and synthesis as described in the paper:
Noise Flow: Noise Modeling with Conditional Normalizing Flows
It also provides code for training and testing a CNN-based image denoiser (DnCNN) using Noise Flow as a noise generator, with comparison to other noise generation methods (i.e., AWGN and signal-dependent noise).
Python (works with 3.6)
TensorFlow (works with 2.0)
TensorFlow Probability (tested with 0.8.0)
Despite not tested, the code may work with library versions other than the specified.
Smartphone Image Denoising Dataset (SIDD)
It is recommended to use the medium-size SIDD for training Noise Flow:
The code checks for and downloads SIDD_Medium_Raw if it does not exist.
Start by running train.ipynb
It contains a set of examples for training different models (as described in the paper) and optionally perform testing and sampling concurrently.
the origianl codes provided by the author have some running problems, and the pretrained model seems to not work well. Therefore, I do some modifications based on the original codes and rewrite the training codes which estimate the algorithm performance using KL between the real noise image and the generating noise image. After a serise of experiment, I find this algorithm is limit for specific cam and iso.