Titre :

Scale-free Texture Segmentation: Expert Feature-based versus Deep Learning strategies

Auteur(s) :

Pascal, Barbara [Auteur]
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
Mauduit, Vincent [Auteur]
Laboratoire de Physique de l'ENS Lyon [Phys-ENS]
Pustelnik, Nelly [Auteur]
Laboratoire de Physique de l'ENS Lyon [Phys-ENS]
Abry, Patrice [Auteur]
Laboratoire de Physique de l'ENS Lyon [Phys-ENS]

Titre de la manifestation scientifique :

28th European Signal Processing Conference

Ville :

Amsterdam

Pays :

Pays-Bas

Date de début de la manifestation scientifique :

2021-01-18

Titre de l’ouvrage :

2020 28th European Signal Processing Conference (EUSIPCO)

Mot(s)-clé(s) en anglais :

Deep learning
CNN
Texture
Segmentation
Fractal
Total variation
Wavelets

Discipline(s) HAL :

Statistiques [stat]/Machine Learning [stat.ML]
Sciences de l'ingénieur [physics]/Traitement du signal et de l'image [eess.SP]
Informatique [cs]/Apprentissage [cs.LG]

Résumé en anglais : [en]

Texture segmentation constitutes a central task in image processing, classically based on two-step procedures consisting first in computing hand-crafted features devised from a priori expert knowledge and second in combining ...
Lire la suite >Texture segmentation constitutes a central task in image processing, classically based on two-step procedures consisting first in computing hand-crafted features devised from a priori expert knowledge and second in combining them into clustering algorithms. Deep learning approaches can be seen as merging these two steps into a single one with both discovering features and performing segmentation. Using fractal textures, often seen as relevant models in real-world applications, the present work compares a recently devised texture segmentation algorithm incorporating expert-driven scale-free features estimation into a Joint TV optimization framework against convolutional neural network architectures. From realistic synthetic textures, comparisons are drawn not only for segmentation performance, but also with respect to computational costs, architecture complexities and robustness against departures between training and testing datasets.Lire moins >

Langue :

Anglais

Comité de lecture :

Oui

Audience :

Internationale

Vulgarisation :

Non

Collections :

• Centre de Recherche en Informatique, Signal et Automatique de Lille (CRIStAL) - UMR 9189

Source :

Harvested from HAL