Tutorial I: Feature Extraction and Denoising: A Saga of U+V Models
I will review the so-called U+V models in image representation and denoising context. Under these models, the original data f is assumed to
consist of the u component that captures the main important "features", and the v component that are mainly textures and noise. Going over the various models, such as Mumford-Shah, Rudin-Osher-Fatemi, DeVore-Lucier, Chen-Donoho-Saunders, Cohen-Daubechies-Dahmen-DeVore, etc., I will discuss how the U component are modeled mathematically, and how much they reflect the reality. I will also discuss the recent interactions between harmonic analysis and PDE for image processing.
This tutorial should naturally leads to the lectures of Donoho and Candes on Friday.