Leo DAVY
PhD Student
@ Ecole Normale Supérieure de Lyon
@ Ecole Normale Supérieure de Lyon
Supervised by Patrice ABRY & Nelly PUSTELNIK
Texture segmentation with proximal unrolled networks
Contact me : leo[.]davy[at]ens-lyon[.]fr
Subject description
In my PhD I'm trying to understand textures in computer vision and devise efficient algorithms that automatically extract those.
- Textures we are interested in are those that relate to the variations of the image. In particular, we identify textures as those parts sharing statistically the same local regularity without assuming isotropy.
- Random fields are relevant models for such textures and we investigate their behaviour under wavelet analysis. We are able to compute similarity metrics on fields like the (Anisotropic)-(Fractional)-Brownian Fields.
- Proximal operator theory allows to solve optimization problems involving several penalizations. Segmentation can be cast as such a problem and we have developped fast proximal algorithms to solve it.
- Unrolled networks are algorithms viewed as neural networks, a natural formulation for our algorithms. This allows us to use tools from statistical machine learning to automatically fine tune our models based on the data.
Publications
International conferences :
With Nelly PUSTELNIK and Patrice ABRY
Accepted at ICASSP 2023 @ Rhodes Island, Greece
National conferences :
With Nelly PUSTELNIK and Patrice ABRY
(Automated unsupervised hyperparameter selection by automatic differentiation, applied to texture segmentation)
Submitted to GRETSI 2023 @ Grenoble
Other
Supervised by Patrice ABRY and Nelly PUSTELNIK @ Laboratoire de Physique, ENSL
Proximal algorithms for anisotropic texture segmentation,
(Masters thesis for M2 Mathématiques Avancées - Statistiques et Probabilités @ Ecole Normale Supérieure, Lyon)
Supervised by Jean-François CROUZET
Sparse reconstruction of signals Frames, Atomic decomposition and introduction to compressed sensing,
@ Institut de Mathématiques Alexander Grothendieck
(Report for M1 Mathématiques Fondamentales @ Faculté des Sciences, Montpellier)
Supervised by Joseph NAJNUDEL @ School of Mathematics, University of Bristol
(Internship report for L3 Mathématiques - Parcours Spécial @ Université Paul Sabatier, Toulouse)
Softwares
Automatic Unsupervised parameter selection for texture segmentation in PyTorch
Undecimating a (dual-tree complex) wavelet transform (Python)
Anisotropic wavelet analysis for textures (Python)
Curriculum
- 2022-2025 PhD with Patrice ABRY and Nelly PUSTELNIK @ Laboratoire de Physique, ENS Lyon
- 2021-2022 Master 2 : Mathématiques Avancées - Parcours Statistiques et Probabilités @ Ecole Normale Supérieure, Lyon
- 2020-2021 Master 1 : Mathématiques Fondamentales @ Faculté des Sciences, Montpellier
- 2019-2020 Master 1 : Energie electrique, Electronique et Automatique - Parcours Robotique @ Faculté des Sciences, Montpellier
- 2016-2019 Licence de Mathématiques - Parcours Spécial @ Université Toulouse 3, Paul Sabatier