Vous êtes ici : Accueil » Kiosque » Annonce

Identification

Identifiant: 
Mot de passe : 

Mot de passe oublié ?
Détails d'identification oubliés ?

Annonce

12 avril 2019

Multivariate Scale Free Texture Analysis


Catégorie : Doctorant


Ph.D Thesis offer

Multivariate Scale-Free Texture Analysis:

Texture segmentation, interface estimation, anomaly detection with applications to real-world images

From multifractal analysis to statistical and deep learning: stability and robustness

 

Starting Date : 01/10/2019

Duration : 3 years

Deadline : 20/04/2019

Funding : DGA Ph.D fellowship (secured)

Requirements : M2 research or equivalent, EU nationality

Place : Ecole Normale Supérieure de Lyon

 

Research team and supervision.

Patrice Abry patrice.abry@ens-lyon.fr http://perso.ens-lyon.fr/patrice.abry/

DR CNRS, Physics Lab., Ecole Normale Supérieure de Lyon

The conduction of the PhD will be framed into a more general research program of the

Signals, Systems and Physics team.

 

Scientific context.

Anomaly detection and texture segmentation are ubiquitous and challenging issues in image

processing and have been envisaged through many different concepts. Over the decades following the

seminal works of Benoît Mandelbrot, an overwhelming number of images have been shown to be

well-characterized by scale-free textures. This implies a major change in paradigm: the

characterization can no longer rely on one specific analysis scale since all scales play equivalent roles,

and the classical signal/image processing tools must thus be replaced with tools that evidence and

quantify the mechanisms relating scales one to the other.

Multifractal analysis has recently matured to become one of the most powerful tools for this purpose.

It benefits from a well-grounded theoretical framework and a robust practical and has been extremely

successful in a large panel of applications of very different natures. Yet, these successes assumed that

i) data are univariate (independent analysis of one image at a time)

ii) data are isotropic (all directions in an image are equivalent)

iii) data are homogeneous (multifractal properties are the same everywhere in the image)

while these requirements are no longer met in many modern real-world applications. Indeed, data are

often naturally multivariate (dependent measurements, captured by different imaging sensors, jointly

convey the information of interest), anisotropic (certain directions in the image have privileged roles)

and often consist of zones, to be detected, whose properties differ from that of the rest of the data.

 

 

Missions and Activities. Research program and targeted contributions

The overarching goal of the proposed Ph.D is to devise and compare texture segmentation andanomaly detection for multivariate, non isotropic and heterogeneous images using tools and concept has recently matured to become one of the most powerful tools for this purpose.

It benefits from a well-grounded theoretical framework and a robust practical and has been extremely

successful in a large panel of applications of very different natures. Yet, these successes assumed that

i) data are univariate (independent analysis of one image at a time)

ii) data are isotropic (all directions in an image are equivalent)

iii) data are homogeneous (multifractal properties are the same everywhere in the image)

while these requirements are no longer met in many modern real-world applications. Indeed, data are

often naturally multivariate (dependent measurements, captured by different imaging sensors, jointly

convey the information of interest), anisotropic (certain directions in the image have privileged roles)

and often consist of zones, to be detected, whose properties differ from that of the rest of the data.

 

 

Missions and Activities. Research program and targeted contributions

The overarching goal of the proposed Ph.D is to devise and compare texture segmentation and

anomaly detection for multivariate, non isotropic and heterogeneous images using tools and concept

ranging from optimization formulation to statistical and deep learning. Emphasis will be on assessing

stability and robustness. Real-world applications stemming from physics, geophysics or biomedical

imaging will be targeted.

The research program is organized around two main directions:

(i) Statistical and deep learning. The PhD student will formally study the link between unsupervised

and supervised strategies for scale-free and multifractal texture segmentations: The former,

already partially developed, relies on estimated multifractal attributes combined into variational

segmentation frameworks, the latter will be based on deep neural networks. The PhD candidate will

first propose variations on state-of-the-art of unsupervised multifractal segmentation strategies to

explore the potential benefits in semi-supervision before turning to self-supervised or fully-supervised

deep neural network strategies. A methodology to conduct relevant and meaningful comparisons

(beyond the simple classificationperformance) between the different approaches will also constitute an

important research track.

(ii) Stability and robustness assessment. Segmentation performance often strongly depend on

hyperparameter tuning. For instance, in the unsupervised fractal current formulation, crucial choices

need to be made by the users (decomposition level or range of scales, properties of the underlying

multiscale quantities, level of redundance in the representation). While current formulations rely on

standard 2D discrete wavelet transform, extension to other multiscale quantities (such as complex

wavelets or shearlets) will be explored. The issue is even more critical in deep learning where the

robustness of achieved results with respect to the network architecture is often debatable. An

important research direction will thus be to devise a sound and robust methodology to assess the

sensitivity of the output to hyperparameter tuning. Providing application experts not only with a

segmentation or classification outcome, but also with confidence levels on both how much the

classification of each pixel/subject should be trusted and on robust the overall classification is (in

terms in number of classes for instance) are key conditions for relevant interpretations of the outcome

with respect to the application.

The method developed during the PhD will be evaluated on both synthetic data and real-world data,

benefiting from the expertise and collaboration network of the SiSyPh team, for a large panel of realworld

applications ranging from geophysics to biomedical data and art investigation.

Application.

All applications must be sent electronically and as soon as possible to P. Abry (minimum: motivation

letter, CV).

 

Only candidates with nationality from EU countries can be considered.

The deadline is 20 April 2019.

 

Selected references.

S.G. Roux, M. Clausel, B. Vedel, S. Jaffard, and P. Abry. “Self-similar anisotropic texture analysis:

the hyperbolic wavelet transform contribution.” IEEE Transaction on Image Processing, 22(11):4353-

4363, 2013

N. Pustelnik, H. Wendt, P. Abry, N. Dobigeon, “Combining Local Regularity Estimation and Total

Variation Optimization for Scale-Free Texture Segmentation.” IEEE Transaction on Computational

Imaging, 2(4): 468-479, Dec. 2016.

 

Dans cette rubrique

(c) GdR 720 ISIS - CNRS - 2011-2018.