Self-similarity based methods in computational imaging

par Andrès Almansa (CNRS - LTCI)

mardi 17 janv. 2017, 15:15 → 16:15 Europe/Paris

4e étage, salle 435 (UMPA, ENS Lyon - Site Monod)

This talk is composed of two parts.
In the first part I give a brief overview of a few problems in computational imaging and why these problems lead to new kinds of ill-posed inverse problems that require novel regularization and statistical estimation approaches to solve them. Then I shall make a brief review of patch-based image restoration methods, and their evolution from simple example-based strategies to the more recent Bayesian approaches based on local Gaussian priors. Without being exhaustive I shall highlight a few trends and open challenges under a common framework. In the second part I shall illustrate these trends with my ongoing work on restoration of noisy images that have been compressed by quantization of their wavelet coefficients. When the noise and quantization levels are within a certain range the resulting image is corrupted by a highly structured noise in the spatial domain. However, in the wavelet domain noise has a particular statistical structure, that can be separated from the image by means of local Gaussian priors. Doing so imposes a number of technical challenges that I will explain how to overcome.