Texture segmentation with wavelets and unrolled proximal networks

Speakers

Leo Davy
Ph.D. Candidate, Physics Laboratory, École Normale Supérieure de Lyon.

Title

"Texture segmentation with wavelets and unrolled proximal networks"

Abstract

Texture segmentation consists in recovering homogeneous parts of an image without relying on its geometry (edges). A stochastic model that allows modeling textures with scale-free anisotropic properties will be considered. These textures can then be analyzed using complex wavelets to recover their features of interest on spatially homogeneous textures. To recover these features locally an inverse regularized problem is proposed and solved by a proximal iterative algorithm leading to an interpretable segmentation method.

Furthermore, the iterative algorithm will be unrolled - i.e. interpreted as a neural network with parameters to be learned - in order to learn regularization parameters using widely available automatic differentiation software. A generic and easy-to-implement restart scheme will then be presented to learn as if the proximal unrolled network were infinitely deep by simplifying the Deep Equilibrium framework using convergence properties of the iterative algorithm. An unsupervised learning setting will also be considered.

About Speakers

Leo Davy is a Ph.D student at the École Normale Supérieure de Lyon (France) where he also obtained his M.Sc. in Probability and Statistics after obtaining a B.S. in Mathematics at Université Toulouse 3. His Ph.D thesis is on the development of interpretable segmentation methods with a focus on multiscale models and is supervised by Patrice Abry and Nelly Pustelnik. His research interests include optimization, wavelet analysis, statistics and stochastic models in image processing.