Croissance de région variationnelle et contraintes géométriques tridimensionnelles pour la segmentation d'image

= Variational region growing and tridimensional shape prior for image segmentation

• Auteur / Autrice : Jean-Loïc Rose
• Direction : Chantal Revol-Muller, Christophe Odet
• Type : Thèse de doctorat
• Discipline(s) : Images et systèmes
• Date : Soutenance en 2008
• Etablissement(s) : Lyon, INSA

Mots clés

Mots clés contrôlés

Tchebychev, Systèmes de

Imagerie médicale

Traitement d'images -- Techniques numériques

Résumé

La segmentation d'image par croissance de région est une méthode permettant de localiser et d'extraire un objet dans une image. Pour répondre aux difficultés de la segmentation, nous nous sommes intéressés à l'intégration d'apriori de forme dans la croissance de région. Deux contributions originales sont présentées dans ce manuscrit. La première contribution consiste à proposer une nouvelle vision de la croissance de région en décrivant son processus dans un cadre variationnel. Dans cette approche, la région est modélisée par une fonction caractéristique discrète. Nous déterminons l'équation d'évolution de cette fonction en adoptant une approche variationnelle. La croissance de région variationnelle se distingue des croissance de région classiques, qui évoluent tant qu'un critère d'homogénéité est respecté, en intégrant une minimisation d'énergie dans son processus d'évolution. Cette stratégie lui assure une convergence, sans avoir à fixer un seuil minimal d'homogénéité. Dans la deuxième contribution, nous définissons une contrainte géométrique tridimensionnelle pour la segmentation, basée sur des descripteurs région. La contrainte de forme proposée est construite à partir des moments de Tchebychev. Un terme de pondération permet d'optimiser notre a priori et de tenir compte de la hiérarchie des moments de Tchebychev. Nous mettons en avant nos contributions en intégrant la contrainte de forme dans la croissance de région variationnelle. L'évaluation qualitative et quantitative de la croissance de région variationnelle est réalisée à partir de tests effectués sur des images synthétiques et biomédicales. Les résultats mettent en évidence l'amélioration apportée par l’a priori de forme et notamment comment celui-ci réussit à contraindre l’évolution de la croissance.

IMAGE segmentation through region growing is a method that localizes and extracts an object from an image. When difficult segmentation context are linked to noise or a lack of information, the use of prior knowledge in the region growing process improves segmentation accuracy. Medical imaging is often concerned by this context. Moreover, medical applications deal with large amounts of data. Then, it is mandatory to use a robust and fast processing. This context leads us to propose a shape prior for the region growing process. The manuscript presents our work composed of two main contributions. Firstly, we proposed to define region growing in the variational framework. We choose to describe our region by a discrete function. The region evolution equation is defined in a variational formulation. An energy functional minimization is associated with our function. For the criterion minimum, the segmented region describes the desired object. The main interest of our region growing approach is the use of region based energy minimization in the evolution process. Then, we define a region-based shape constrain in the segmentation. We define shape descriptors based on Tchebychev moments. We proposed to introduce a weighting term into our functional. This term allows to consider the Tchebychev moments hierarchy, thus adapting our criterion with the encoded moments information. We have proposed a shape prior which incorporates invariance with respect to the group of similarity transformations of a given shape. Our shape constrain is defined for two-dimensional and three-dimensional images. We apply our shape prior in the variational region growing. This application shows the ability of our method to incorporate region-based energy. We propose to study the ability of our shape constrain to influence the evolution of the region growing near the desired shape. Segmentation results are given for synthetic and real images. IMAGE segmentation through region growing is a method that localizes and extracts an object from an image. When difficult segmentation context are linked to noise or a lack of information, the use of prior knowledge in the region growing process improves segmentation accuracy. Medical imaging is often concerned by this context. Moreover, medical applications deal with large amounts of data. Then, it is mandatory to use a robust and fast processing. This context leads us to propose a shape prior for the region growing process. The manuscript presents our work composed of two main contributions. Firstly, we proposed to define region growing in the variational framework. We choose to describe our region by a discrete function. The region evolution equation is defined in a variational formulation. An energy functional minimization is associated with our function. For the criterion minimum, the segmented region describes the desired object. The main interest of our region growing approach is the use of region based energy minimization in the evolution process. Then, we define a region-based shape constrain in the segmentation. We define shape descriptors based on Tchebychev moments. We proposed to introduce a weighting term into our functional. This term allows to consider the Tchebychev moments hierarchy, thus adapting our criterion with the encoded moments information. We have proposed a shape prior which incorporates invariance with respect to the group of similarity transformations of a given shape. Our shape constrain is defined for two-dimensional and three-dimensional images. We apply our shape prior in the variational region growing. This application shows the ability of our method to incorporate region-based energy. We propose to study the ability of our shape constrain to influence the evolution of the region growing near the desired shape. Segmentation results are given for synthetic and real images.