Practical Noise Generator

A practical noise generator is provided in this repository.

Noise Generator:

The Poissonian-Gaussian model is used to the model the noise. The Poissonian-Gaussian model model has two parameters: 𝑎 and 𝑏, where 𝑎 is related to the variance of Poissonian component and 𝑏 the variance of the Gaussian component. The Smartphone Image Denoising Dataset (SIDD) is used to estimate 𝑎 and 𝑏 for each of the color components (R,G,B).

The following example generates N number of noisy image for the given $IMG_DIR using the provided noise generator and saves them in OUT_DIR

python ./noise_sampling.py --img_dir $IMG_DIR --n_obs $N --out_dir $OUT_DIR

The code first obtains (𝑎, 𝑏) for each color component. Then generates noisy image according to the sampled parameters. img_syn_noisy_q variable is the quantized noisy image that can be stored or used in the training dataloader.

How to set up

1. create conda environment: conda create -n test_generator python=3.6.9
2. activate environment: conda activate test_generator
3. install dependencies: pip install -r requirements.txt