• Calendar

January 10, Tuesday
12:00 – 13:00

Many image restoration tasks are ill posed problems, often solved with image priors. Since image priors are only approximate, in general this yields suboptimal restoration results. Given the large body of work on image priors and on image restoration algorithms, this begs fundamental questions of what are optimal restoration procedures, what are (if any) inherent limitations imposed by the statistics of natural images, and what potential gains we may expect from additional years of research efforts.

In this talk we focus on these problems for the simplest restoration task of natural image denoising, where the goal is to estimate a clean natural image, given a noise-corrupted version of it.

We propose a statistical framework and a non-parametric computational approach to study these questions: what is optimal natural image denoising ? what are its fundamental lower bounds, and how far are current algorithms from optimality ?

As we shall see, answers to these questions involve both computational limitations, information-statistical issues and a fundamental property of natural images - scale invariance. Their combination allows us to give a ballpark estimate on the best achievable denoising, and to suggest directions for potential improvements of current algorithms.

Joint work with Anat Levin, Fredo Durand and Bill Freeman.