SEMINAIRE DU 5 FÉVRIER 2015 – 16H @ MAP5 – SALLE DU CONSEIL
From Theory to Practice: a Tour of Image Denoising
Denoising is one of the most crucial issue for the quality of numerical images, and it has to be done before anything else to obtain optimal quality.
In the case of raw images, the noise can be model by a Poisson noise, and therefore a variance stabilizing transform can be applied to the image to go back to the ideal case of white Gaussian noise.
In this simple case, a lot of algorithms have been proposed, and by analyzing them and their generic tools (aggregation, color space transform, oracle iteration, …) we are able to propose a new one: the NL-Bayes algorithm.
As real-life images are not raw images but JPEG ones, this simple model cannot be applied directly as it. Therefore it needs to be adapted to signal-dependent noise.
By adapting the NL-Bayes algorithm to signal-dependent noise, coupling it with a Noise estimation algorithm and applying a multi-scale approach, we are able to propose an efficient blind denoising algorithm, the Noise Clinic.
Marc Lebrun, From Theory to Practice, A Tour of Image Denoising, PhD Thesis CMLA, ENS Cachan,12 juin 2014. [pdf]