HDR : Representation methods for image analysis and interpretations

Accreditation to supervise research: Naoufel Werghi

Title: Representation methods for image analysis and interpretations

Abstract: This thesis exposes computer vision research works covering different aspects of image analysis and interpretation. In this context, a large part of my contributions has been devoted to the representation. This concept can be described as the development of an optimal computational framework for encoding visual data to ensure an effective solution for given computer vision problem.

• The first part of the presentation addresses the paradigm of shape representation in images of articulated and tubular objects that come in the form of a scattered plot. We will discuss solutions based on a Reeb-graph diagram, we will describe also the mechanisms proposed to solve problems emanating from the variability of posture and noise that often shatters this category of visual data.
• The second part will be devoted to the design of an appropriate representations for the analysis of triangular meshes. In this framework, methods for generating ordered structures on triangular meshes, which can be deployed locally and globally will be presented. We will also demonstrate the use of these structures in face surface analysis, and the extension of local binary models to triangular mesh. Then we will present application of this new representation to 3D face recognition and recognition of relief images.
• The third part will focus on a contribution related to the field of medical imaging, always within the framework of representation, we propose here a computer-assisted diagnostic system to assess the severity of an ocular pathology called posterior opacification of the capsule.
• The fourth part will be dedicated to the concept of representation in depth learning perspective. The recent advent of deep learning paradigms was an opportunity to explore approaches to define and extract appropriate representations through deep learning systems. In this context, two applications will be presented dealing with: 1) face recognition in extreme poses and 2) polyp detection in colonoscopy images.
• The presentation closes with research work prospects in the near and medium future.

The examination board is composed of : M. Boulbaba Ben Amor, Professeur des universités, IMT Lille Douai (Rapporteur), M. Jean-Luc Dugelay, Professeur des universités, EURECOM (Rapporteur), M. Fabrice Mériaudeau, Professeur des universités, Université de Bourgogne (Rapporteur), M. Ernest Hirsch, Professeur des universités, Université de Strasbourg (Examinateur), et M. Christophe Doignon, Professeur des universités, Université de Strasbourg (Garant)

The presentation will take place Wednesday 27 June 2018, 9h00 in the amphitheater 207 at Télécom Physique Strasbourg (Illkirch).