• Noise level. The noise level specifies the standard deviation of the noise relative to the intensity range [0,255].

• Noise model. The noise model specifies the distribution of the noise to be Gaussian, Laplace, or Poisson. Gaussian is a good choice in general. Laplace is better suited for heavy tail distributions like salt-and-pepper and dark shot noise. Poisson models low-light acquisition and roughly approximates multiplicative noise.

• The algorithm can run in two different ways:

1. “Denoise” directly denoises the selected image. No noise is added to the image. The noise level specifies the standard deviation of the noise present in the input image.
2. With “Add noise, then denoise,” noise is simulated, then the algorithm is applied to denoise the noisy image.

• Source code for this demo: zip tar.gz