A esfera é utilizada como domínio computacional na modelagem de diversos fenômenos físicos, como em previsão numérica do tempo. Sua discretização pode ser feita de diversas formas, sendo comum o uso de malha regulares em latitude/longitude. Recentemente, também para melhor uso de computação paralela, há uma tendência ao uso de malhas mais isotrópicas, dentre as quais a icosaédrica. Apesar de já existirem modelos atmosféricos que usam malhas icosaédricas, não há consenso sobre as metodologias mais adequadas a esse tipo de malha. Nos propusemos, portanto, a estudar em detalhe diversos fatores envolvidos no desenvolvimento de modelos atmosféricos globais usando malhas geodésicas icosaédricas. A discretização usual por volumes finitos para divergente de um campo vetorial utiliza como base o Teorema da Divergência e a regra do ponto médio nas arestas das células computacionais. A distribuição do erro obtida com esse método apresenta uma forte relação com características geométricas da malha. Definimos o conceito geométrico de alinhamento de células computacionais e desenvolvemos uma teoria que serve de base para explicar interferências de malha na discretização usual do divergente. Destacamos os impactos de certas relações de alinhamento das células na ordem da discretização do método. A teoria desenvolvida se aplica a qualquer malha geodésica e também pode ser usada para os operadores rotacional e laplaciano. Investigamos diversos métodos de interpolação na esfera adequados a malhas icosaédricas, e abordamos o problema de interpolação e reconstrução vetorial na esfera em malhas deslocadas. Usamos métodos alternativos de reconstrução vetorial aos usados na literatura, em particular, desenvolvemos um método híbrido de baixo custo e boa precisão. Por fim, utilizamos as técnicas de discretização, interpolação e reconstrução vetorial analisadas em um método semi-lagrangiano para o transporte na esfera em malhas geodésicas icosaédricas. Realizamos experimentos computacionais de transporte, incluindo testes de deformações na distribuição do campo transportado, que mostraram a adequação da metodologia para uso em modelos atmosféricos globais. A plataforma computacional desenvolvida nesta tese, incluindo geração de malhas, interpolações, reconstruções vetoriais e um modelo de transporte, fornece uma base para o futuro desenvolvimento de um modelo atmosférico global em malhas icosaédricas. / Spherical domains are used to model many physical phenomena, as, for instance, global numerical weather prediction. The sphere can be discretized in several ways, as for example a regular latitude-longitude grid. Recently, also motivated by a better use of parallel computers, more isotropic grids have been adopted in atmospheric global circulation models. Among those, the icosahedral grids are promising. Which kind of discretization methods and interpolation schemes are the best to use on those grids are still a research subject. Discretization of the sphere may be done in many ways and, recently, to make better use of computational resources, researchers are adopting more isotropic grids, such as the icosahedral one. In this thesis, we investigate in detail the numerical methodology to be used in atmospheric models on icosahedral grids. The usual finite volume method of discretization of the divergence of a vector field is based on the divergence theorem and makes use of the midpoint rule for integration on the edges of computational cells. The error distribution obtained with this method usually presents a strong correlation with some characteristics of the icosahedral grid. We introduced the concept of cell alignment and developed a theory which explains the grid imprinting patterns observed with the usual divergence discretization. We show how grid alignment impacts in the order of the divergence discretization. The theory developed applies to any geodesic grid and can also be used for other operators such as curl and Laplacian. Several interpolation schemes suitable for icosahedral grids were analysed, including the vector interpolation and reconstruction problem on staggered grids. We considered alternative vector reconstruction methods, in particular, we developed a hybrid low cost and good precision method. Finally, employing the discretizations and interpolations previously analysed, we developed a semi-Lagrangian transport method for geodesic icosahedral grids. Several tests were carried out, including deformational test cases, which demonstrated that the methodology is suitable to use in global atmospheric models. The computational platform developed in this thesis, including mesh generation, interpolation, vector reconstruction and the transport model, provides a basis for future development of global atmospheric models on icosahedral grids.

Ce travail étudie les variétés pseudo-riemanniennes compactes localement homogènes à travers le prisme des (G,X)-structures, introduites par Thurston dans son programme de géométrisation. Nous commençons par présenter la problématique générale et discutons notamment du rapport entre la complétude géodésique de ces variétés et une autre notion de complétude propre aux (G,X)-structures. Nous donnons également dans le chapitre 1 une nouvelle preuve d’un théorème de Bromberg et Medina qui classifie les métriques lorentziennes invariantes à gauche sur SL(2,R) dont le flot géodésique est complet. Conjecturalement, toute (G,X)-structure pseudo-riemannienne sur une variété compacte est complète. Nous prouvons ici que cela est vrai pour certaines géométries, sous l’hypothèse que la (G,X)-structure est a priori kleinienne. On en déduit que, pour ces géométries, la complétude est une condition fermée. Lorsque X est un groupe de Lie de rang 1 muni de sa métrique de Killing, ce résultat complète un théorème de Guéritaud–Guichard–Kassel–Wienhard selon lequel la complétude est une condition ouverte. Nous nous tournons ensuite vers l’étude des représentations d’un groupe de surface à valeurs dans les isométries d’une variété riemannienne M complète simplement connexe de courbure sectionnelle inférieure à -1. Étant donnée une telle représentation ρ, nous montrons que l’ensemble des représentations fuchsiennes j telles qu’il existe une application (j,ρ)-équivariante et contractante de H2 dans M est un ouvert non vide et contractile de l’espace de Teichmüller (sauf lorsque ρ est elle-même fuchsienne). Ce résultat nous permet de décrire l’espace des métriques lorentziennes de courbure constante -1 sur un fibré en cercle au-dessus d’une surface compacte. Nous montrons que cet espace possède un nombre fini de composantes connexes classifiées par un invariant que nous appelons longueur de la fibre. Nous prouvons également que le volume total de ces métriques ne dépend que de la topologie du fibré et de la longueur de la fibre. / In this work, we study closed locally homogeneous pseudo-Riemannian manifolds through the notion of (G,X)-structure, introduced by Thurston in his geometrization program. We start by presenting the general problem. In particular, we discuss the link between geodesical completeness of those manifolds and another notion of completeness specific to (G,X)-structures. In chapter 1, we also give a new proof of a theorem by Bromberg and Medina which classifies left invariant Lorentz metrics on SL(2,R) that are geodesically complete. Conjecturally, every pseudo-riemannian (G,X)-structure on a closed manifold is complete. Here we prove that it holds for certain geometries, provided that the (G,X )-structure is a priori Kleinian . This implies that, for such geometries, completeness is a closed condition. When X is a Lie group of rank 1 handled with its Killing metric, this result complements a theorem of Guéritaud–Guichard–Kassel–Wienhard, acording to which completeness is an open condition. We then turn to the study of representations of surface groups into the isometry group of a complete simply connected Riemannian manifold M of curvature less than or equal to -1. Given such a representation ρ, we prove that the set of Fuchsian representations j for which there exists a (j,ρ)-equivariant contracting map from H2 to M is a non-empty open contractible subset of the Teichmüller space (unless ρ itself is Fuchsian). This result allows us to describe the space of Lorentz metrics of constant curvature -1 on a circle bundle over a closed surface. We show that this space has finitely many connected components, classified by an invariant that we call the length of the fiber. We also prove that the total volume of those metrics only depends on the topology of the bundle and on the length of the fiber.

In the sub-THz frequency range, the geodesic lens can realize low losses and beam scanning capability with high gain and high aperture efficiency due to its fully metallic property and rotational symmetry. Therefore, in high-frequency applications, a geodesic lens is considered a more promising solution in comparison to phased arrays or other beamforming techniques. To realize dual polarization for geodesic lenses, a polarization rotator using fully metallic screens can be placed at the lens aperture to increase the channel capacity. In this thesis, we propose a dual-polarized fully metallic geodesic lens antenna with the operation frequency centered at 120 GHz. The proposed design contains two layers of geodesic lenses and two polarization rotators placed in their respective apertures with ±45° polarization. By using a twist waveguide for the feeding, we eliminate the leakage caused by the air gap between the metal plates. The simulation results show that the dual-polarized lens can achieve an angular scanning range of ±60° and its scanning loss is 0.6 dB, with an aperture efficiency of 90%. Finally, we propose a prototype design with mechanical considerations to ensure robustness in future manufacturing, assembly, and testing. / Geodetiska linser kan tillämpas åt frekvenser under THz för att realisera låga förluster och strålscanningskapacitet med både hög förstärkning och apertureffektivitet, detta på grund av dess fullt metalliska egenskaper samt rotationssymmetri. I högfrekventa tillämpningar anses därför en geodetisk lins vara lovande i jämförelse med en fasstyrda gruppantenn eller andra strålformningstekniker. Ytterligare så kan polarisationsrotatorer med helt metalliska skärmar kan placeras vid linsöppningen för att realisera korspolariserade fält samt öka kanalkapaciteten hos geodetiska linser. I denna avhandling föreslårs en justerbar korspolariserad samt fullt metallisk geodetisk linsantenn centrerad runt 120 GHz. Den föreslagna designen innehåller två lager geodetiska linser och två polarisationsrotatorer placerade i sina respektive utgångar med respektive polarisatonsförskjutning på ±45°. Genom att använda en vriden vågledare för matningen så eliminerars läckaget som normalt följer av luftgapet mellan metallplattorna. Simuleringsresultaten visar att den korspolariserade linsen kan uppnå ett avsökningsområde inom vinklar ±60° men en skanningsförlust på 0,6 dB, detta med en apertureffektivitet på 90%. Slutligen föreslår vi en prototyp med hänsyn till mekaniska aspekter för att säkerställa robusthet i framtida tillverkning, montering och testning.

Segmentar uma imagem consiste em particioná-la em regiões relevantes para uma dada aplicação, como para isolar um objeto de interesse no domínio de uma imagem. A segmentação é um dos problemas mais fundamentais e desafiadores em processamento de imagem e visão computacional. Ela tem desempenhado um papel importante, por exemplo, na pesquisa em neurologia, envolvendo imagens de Ressonância Magnética (RM), para fins de diagnóstico e tratamento de doenças relacionadas com alterações na anatomia do cérebro humano. Métodos de segmentação baseados na transformada imagem- floresta (IFT, Image Foresting Transform), com funções de conexidade suaves, possuem resultados ótimos, segundo o critério da otimalidade dos caminhos descrito no artigo original da IFT, e têm sido usados com sucesso em várias aplicações, como por exemplo na segmentação de imagens RM de 1.5 Tesla. No entanto, esses métodos carecem de restrições de regularização de borda, podendo gerar segmentações com fronteiras muito irregulares e indesejadas. Eles também não distinguem bem entre bordas similares com orientações opostas, e possuem alta sensibilidade à estimativa dos pesos das arestas do grafo, gerando problemas em imagens com efeitos de inomogeneidade. Nesse trabalho são propostas extensões da IFT, do ponto de vista teórico e experimental, através do uso de funções de conexidade não suaves, para a segmentação interativa de imagens por região. A otimalidade dos novos métodos é suportada pela maximização de energias de corte em grafo, ou como o fruto de uma sequência de iterações de otimização de caminhos em grafos residuais. Como resultados principais temos: O projeto de funções de conexidade mais adaptativas e flexíveis, com o uso de pesos dinâmicos, que permitem um melhor tratamento de imagens com forte inomogeneidade. O uso de grafos direcionados, de modo a explorar a polaridade de borda dos objetos na segmentação por região, e o uso de restrições de forma que ajudam a regularizar a fronteira delineada, favorecendo a segmentação de objetos com formas mais regulares. Esses avanços só foram possíveis devido ao uso de funções não suaves. Portanto, a principal contribuição desse trabalho consiste no suporte teórico para o uso de funções não suaves, até então evitadas na literatura, abrindo novas perpectivas na pesquisa de processamento de imagens usando grafos. / Segmenting an image consist in to partition it into relevant regions for a given application, as to isolate an object of interest in the domain of an image. Segmentation is one of the most fundamental and challenging problems in image processing and computer vision. It has played an important role, for example, in neurology research, involving images of Magnetic Resonance (MR), for the purposes of diagnosis and treatment of diseases related to changes in the anatomy of the human brain. Segmentation methods based on the Image Foresting Transform (IFT), with smooth connectivity functions, have optimum results, according to the criterion of path optimality described in the original IFT paper, and have been successfully used in many applications as, for example, the segmentation of MR images of 1.5 Tesla. However, these methods present a lack of boundary regularization constraints and may produce segmentations with quite irregular and undesired boundaries. They also do not distinguish well between similar boundaries with opposite orientations, and have high sensitivity to the arc-weight estimation of the graph, producing poor results in images with strong inhomogeneity effects. In this work, we propose extensions of the IFT framework, from the theoretical and experimental points of view, through the use of non-smooth connectivity functions for region-based interactive image segmentation. The optimality of the new methods is supported by the maximization of graph cut energies, or as the result of a sequence of paths optimizations in residual graphs. We have as main results: The design of more adaptive and flexible connectivity functions, with the use of dynamic weights, that allow better handling of images with strong inhomogeneity. The use of directed graphs to exploit the boundary polarity of the objects in region-based segmentation, and the use of shape constraints that help to regularize the segmentation boundary, by favoring the segmentation of objects with more regular shapes. These advances were only made possible by the use of non-smooth functions. Therefore, the main contribution of this work is the theoretical support for the usage of non-smooth functions, which were until now avoided in literature, opening new perspectives in the research of image processing using graphs.

Cette thèse traite de l’étude des objets reliés au codage de Bowen-Series du flot géodésiquepour des surfaces hyperboliques de volume fini. On démontre d’abord que le billard géodésiqueassocié à domaine fondamental even corners d’un groupe fuchsien cofini est conjuguéà une bijection du tore, appelée codage étendu, dont l’un des facteurs est la transformationde Bowen-Series. L’intérêt principal de cette conjugaison est qu’elle ne fait toujours intervenirqu’un nombre fini d’objets. On retrouve ensuite des résultats classiques sur le codage deBowen-Series : il est orbite-équivalent au groupe, ses points périodiques sont denses, et ses orbitespériodiques sont en bijection avec les classes d’équivalence d’hyperboliques primitifs dugroupe ; ce qui permet finalement de relier sa fonction zeta de Ruelle à la fonction zeta de Selberg.Les preuves de ces résultats s’appuient sur un lemme combinatoire qui abstrait la propriétéd’orbite-équivalence à des familles de relations qui peuvent être définies sur tout ensemble surlequel agit le groupe. Il est aussi possible de conjuguer le codage étendu à un sous-shift detype fini, sauf pour un ensemble dénombrable de points. Enfin, on prouve que les distributionspropres pour la valeur propre 1 de l’opérateur de transfert sont les distributions de Helgason defonctions propres du laplacien sur la surface, puis que l’on peut associer à toute telle distributionpropre une fonction propre non triviale de l’opérateur de transfert et que ce procédé admet uninverse dans certains cas. / This thesis focuses on the study of the objects linked to the Bowen-Series coding of the geodesicflow for hyperbolic surfaces of finite volume. It is first proved that the geodesic billiardassociated with an even corners fundamental domain for a cofinite fuchsian group is conjugatedwith a bijection of the torus, called extended coding, one factor of which is the Bowen-Seriestransform. The sharpest property of that conjugacy is that it always only involves a finite numberof objects. Some classical results about the Bowen-Series coding are then rediscovered : itis orbit-equivalent with the group, its periodic points are dense, and its periodic orbits are inbijection with conjugacy classes of primitive hyperbolic isometries ; which eventually links itsRuelle zeta function to the Selberg zeta function. The proofs of those results use a combinatoriallemma that abstracts the orbit-equivalence property to families of relations that can be definedon every set on which the group acts. The extended coding is also proved to be conjugated witha subshift of finite type, except for a countable set of points. Finally, it is shown that eigendistributionsof the transfer operator for the eigenvalue 1 are the Helgason boundary values ofeigenfunction of laplacian on the surface, plus that one can associate to each such eigendistributiona non-trivial eigenfunction of the transfer operator and that this process has a reciprocalin some cases.

Dans tout ce qui suit, nous considérons une variété riemannienne compacte de dimension au moins égale à 2. A tout domaine (suffisamment régulier) , on peut associer la première valeur propre ?Ù de l’opérateur de Laplace-Beltrami avec condition de Dirichlet au bord. Nous dirons qu’un domaine est extrémal (sous entendu, pour la première valeur propre de l’opérateur de Laplace-Beltrami) si est un point critique de la fonctionnelle Ù? ?O sous une contrainte de volume V ol(Ù) = c0. Autrement dit, est extrémal si, pour toute famille régulière {Ot}te (-t0,t0) de domaines de volume constant, telle que Ù 0 = Ù, la dérivée de la fonction t ? ?Ot en 0 est nulle. Rappelons que les domaines extrémaux sont caractérisés par le fait que la fonction propre, associée à la première valeur propre sur le domaine avec condition de Dirichlet au bord, a une donnée de Neumann constante au bord. Ce résultat a été démontré par A. El Soufi et S. Ilias en 2007. Les domaines extrémaux sont donc des domaines sur lesquels peut être résolu un problème elliptique surdéterminé. L’objectif principal de cette thèse est la construction de domaines extrémaux pour la première valeur propre de l’opérateur de Laplace-Beltrami avec condition de Dirichlet au bord. Nous donnons des résultats d’existence de domaines extrémaux dans le cas de petits volumes ou bien dans le cas de volumes proches du volume de la variété. Nos résultats permettent ainsi de donner de nouveaux exemples non triviaux de domaines extrémaux. Le premier résultat que nous avons obtenu affirme que si une variété admet un point critique non dégénéré de la courbure scalaire, alors pour tout volume petit il existe un domaine extrémal qui peut être construit en perturbant une boule géodésique centrée en ce point critique non dégénéré de la courbure scalaire. La méthode que nous utilisons pour construire ces domaines extrémaux revient à étudier l’opérateur (non linéaire) qui à un domaine associe la donnée de Neumann de la première fonction propre de l’opérateur de Laplace-Beltrami sur le domaine. Il s’agit d’un opérateur (hautement non linéaire), nonlocal, elliptique d’ordre 1. Dans Rn × R/Z, le domaine cylindrique Br × R/Z, o`u Br est la boule de rayon r > 0 dans Rn, est un domaine extrémal. En étudiant le linéarisé de l’opérateur elliptique du premier ordre défini par le problème précédent et en utilisant un résultat de bifurcation, nous avons démontré l’existence de domaines extrémaux nontriviaux dans Rn × R/Z. Ces nouveaux domaines extrémaux sont proches de domaines cylindriques Br × R/Z. S’ils sont invariants par rotation autour de l’axe vertical, ces domaines ne sont plus invariants par translations verticales. Ce deuxi`eme r´esultat donne un contre-exemple à une conjecture de Berestycki, Caffarelli et Nirenberg énoncée en 1997. Pour de grands volumes la construction de domaines extrémaux est techniquement plus difficile et fait apparaître des phénomènes nouveaux. Dans ce cadre, nous avons dû distinguer deux cas selon que la première fonction propre Ø0 de l’opérateur de Laplace-Beltrami sur la variété est constante ou non. Les résultats que nous avons obtenus sont les suivants : 1. Si Ø0 a des points critiques non dégénérés (donc en particulier n’est pas constante), alors pour tout volume assez proche du volume de la variété, il existe un domaine extrémal obtenu en perturbant le complément d’une boule géodésique centrée en un des points critiques non dégénérés de Ø0. 2. Si Ø0 est constante et la variété admet des points critiques non dégénérés de la courbure scalaire, alors pour tout volume assez proche du volume de la variété il existe un domaine extrémal obtenu en perturbant le complément d’une boule géodésique centrée en un des points critiques non dégénérés de la courbure scalaire / In what follows, we will consider a compact Riemannian manifold whose dimension is at least 2. Let Ù be a (smooth enough) domain and ?O the first eigenvalue of the Laplace-Beltrami operator on Ù with 0 Dirichlet boundary condition. We say that Ù is extremal (for the first eigenvalue of the Laplace-Beltrami operator) if is a critical point for the functional Ù? ?O with respect to variations of the domain which preserve its volume. In other words, Ù is extremal if, for all smooth family of domains { Ù t}te(-t0,t0) whose volume is equal to a constant c0, and Ù 0 = Ù, the derivative of the function t ? ?Ot computed at t = 0 is equal to 0. We recall that an extremal domain is characterized by the fact that the eigenfunction associated to the first eigenvalue of the Laplace-Beltrami operator over the domain with 0 Dirichlet boundary condition, has constant Neumann data at the boundary. This result has been proved by A. El Soufi and S. Ilias in 2007. Extremal domains are then domains over which can be solved an elliptic overdeterminated problem. The main aim of this thesis is the construction of extremal domains for the first eigenvalue of the Laplace-Beltrami operator with 0 Dirichlet boundary condition. We give some existence results of extremal domains in the cases of small volume or volume closed to the volume of the manifold. Our results allow also to construct some new nontrivial exemples of extremal domains. The first result we obtained states that if the manifold has a nondegenerate critical point of the scalar curvature, then, given a fixed volume small enough, there exists an extremal domain that can be constructed by perturbation of a geodesic ball centered in that nondegenerated critical point of the scalar curvature. The methode used is based on the study of the operator that to a given domain associes the Neumann data of the first eigenfunction of the Laplace-Beltrami operator over the domain. It is a highly nonlinear, non local, elliptic first order operator. In Rn × R/Z, the circular-cylinder-type domain Br × R/Z, where Br is the ball of radius r > 0 in Rn, is an extremal domain. By studying the linearized of the elliptic first order operator defined in the previous problem, and using some bifurcation results, we prove the existence of nontrivial extremal domains in Rn × R/Z. Such extremal domains are closed to the circular-cylinder-type domains Br × R/Z. If they are invariant by rotation with respect to the vertical axe, they are not invariant by vertical translations. This second result gives a counterexemple to a conjecture of Berestycki, Caffarelli and Nirenberg stated in 1997. For big volumes the construction of extremal domains is technically more difficult and shows some new phenomena. In this context, we had to distinguish two cases, according to the fact that the first eigenfunction Ø0 of the Laplace-Beltrami operator over the manifold is constant or not. The results obtained are the following : 1. If Ø0 has a nondegenerated critical point (in particular it is not constant), then, given a fixed volume closed to the volume of the manifold, there exists an extremal domain obtained by perturbation of the complement of a geodesic ball centered in a nondegenerated critical point of Ø0. 2. If Ø0 is constant and the manifold has some nondegenerate critical points of the scalar curvature, then, for a given fixed volume closed to the volume of the manifold, there exists an extremal domain obtained by perturbation of the complement of a geodesic ball centered in a nondegenerate critical point of the scalar curvature

Segmentar uma imagem consiste em particioná-la em regiões relevantes para uma dada aplicação, como para isolar um objeto de interesse no domínio de uma imagem. A segmentação é um dos problemas mais fundamentais e desafiadores em processamento de imagem e visão computacional. Ela tem desempenhado um papel importante, por exemplo, na pesquisa em neurologia, envolvendo imagens de Ressonância Magnética (RM), para fins de diagnóstico e tratamento de doenças relacionadas com alterações na anatomia do cérebro humano. Métodos de segmentação baseados na transformada imagem- floresta (IFT, Image Foresting Transform), com funções de conexidade suaves, possuem resultados ótimos, segundo o critério da otimalidade dos caminhos descrito no artigo original da IFT, e têm sido usados com sucesso em várias aplicações, como por exemplo na segmentação de imagens RM de 1.5 Tesla. No entanto, esses métodos carecem de restrições de regularização de borda, podendo gerar segmentações com fronteiras muito irregulares e indesejadas. Eles também não distinguem bem entre bordas similares com orientações opostas, e possuem alta sensibilidade à estimativa dos pesos das arestas do grafo, gerando problemas em imagens com efeitos de inomogeneidade. Nesse trabalho são propostas extensões da IFT, do ponto de vista teórico e experimental, através do uso de funções de conexidade não suaves, para a segmentação interativa de imagens por região. A otimalidade dos novos métodos é suportada pela maximização de energias de corte em grafo, ou como o fruto de uma sequência de iterações de otimização de caminhos em grafos residuais. Como resultados principais temos: O projeto de funções de conexidade mais adaptativas e flexíveis, com o uso de pesos dinâmicos, que permitem um melhor tratamento de imagens com forte inomogeneidade. O uso de grafos direcionados, de modo a explorar a polaridade de borda dos objetos na segmentação por região, e o uso de restrições de forma que ajudam a regularizar a fronteira delineada, favorecendo a segmentação de objetos com formas mais regulares. Esses avanços só foram possíveis devido ao uso de funções não suaves. Portanto, a principal contribuição desse trabalho consiste no suporte teórico para o uso de funções não suaves, até então evitadas na literatura, abrindo novas perpectivas na pesquisa de processamento de imagens usando grafos. / Segmenting an image consist in to partition it into relevant regions for a given application, as to isolate an object of interest in the domain of an image. Segmentation is one of the most fundamental and challenging problems in image processing and computer vision. It has played an important role, for example, in neurology research, involving images of Magnetic Resonance (MR), for the purposes of diagnosis and treatment of diseases related to changes in the anatomy of the human brain. Segmentation methods based on the Image Foresting Transform (IFT), with smooth connectivity functions, have optimum results, according to the criterion of path optimality described in the original IFT paper, and have been successfully used in many applications as, for example, the segmentation of MR images of 1.5 Tesla. However, these methods present a lack of boundary regularization constraints and may produce segmentations with quite irregular and undesired boundaries. They also do not distinguish well between similar boundaries with opposite orientations, and have high sensitivity to the arc-weight estimation of the graph, producing poor results in images with strong inhomogeneity effects. In this work, we propose extensions of the IFT framework, from the theoretical and experimental points of view, through the use of non-smooth connectivity functions for region-based interactive image segmentation. The optimality of the new methods is supported by the maximization of graph cut energies, or as the result of a sequence of paths optimizations in residual graphs. We have as main results: The design of more adaptive and flexible connectivity functions, with the use of dynamic weights, that allow better handling of images with strong inhomogeneity. The use of directed graphs to exploit the boundary polarity of the objects in region-based segmentation, and the use of shape constraints that help to regularize the segmentation boundary, by favoring the segmentation of objects with more regular shapes. These advances were only made possible by the use of non-smooth functions. Therefore, the main contribution of this work is the theoretical support for the usage of non-smooth functions, which were until now avoided in literature, opening new perspectives in the research of image processing using graphs.

Dans les domaines de l’imagerie médicale et de la vision par ordinateur, la segmentation joue un rôle crucial dans le but d’extraire les composantes intéressantes d’une image ou d’une séquence d’images. Elle est à l’intermédiaire entre le traitement d’images de bas niveau et les applications cliniques et celles de la vision par ordinateur de haut niveau. Ces applications de haut niveau peuvent inclure le diagnostic, la planification de la thérapie, la détection et la reconnaissance d'objet, etc. Parmi les méthodes de segmentation existantes, les courbes géodésiques minimales possèdent des avantages théoriques et pratiques importants tels que le minimum global de l’énergie géodésique et la méthode bien connue de Fast Marching pour obtenir une solution numérique. Dans cette thèse, nous nous concentrons sur les méthodes géodésiques basées sur l’équation aux dérivées partielles, l’équation Eikonale, afin d’étudier des méthodes précises, rapides et robustes, pour l’extraction de structures tubulaires et la segmentation d’image, en développant diverses métriques géodésiques locales pour des applications cliniques et la segmentation d’images en général. / In the fields of medical imaging and computer vision, segmentation plays a crucial role with the goal of separating the interesting components from one image or a sequence of image frames. It bridges the gaps between the low-level image processing and high level clinical and computer vision applications. Among the existing segmentation methods, minimal geodesics have important theoretical and practical advantages such as the global minimum of the geodesic energy and the well-established fast marching method for numerical solution. In this thesis, we focus on the Eikonal partial differential equation based geodesic methods to investigate accurate, fast and robust tubular structure extraction and image segmentation methods, by developing various local geodesic metrics for types of clinical and segmentation tasks. This thesis aims to applying different geodesic metrics based on the Eikonal framework to solve different image segmentation problems especially for tubularity segmentation and region-based active contours models, by making use of more information from the image feature and prior clinical knowledges. The designed geodesic metrics basically take advantages of the geodesic orientation or anisotropy, the image feature consistency, the geodesic curvature and the geodesic asymmetry property to deal with various difficulties suffered by the classical minimal geodesic models and the active contours models. The main contributions of this thesis lie at the deep study of the various geodesic metrics and their applications in medical imaging and image segmentation. Experiments on medical images and nature images show the effectiveness of the presented contributions.

L'objet de cette thèse est d'étendre le cadre théorique et les méthodes numériques du transport optimal à des objets plus généraux que des mesures de probabilité. En premier lieu, nous définissons des modèles de transport optimal entre mesures positives suivant deux approches, interpolation et couplage de mesures, dont nous montrons l'équivalence. De ces modèles découle une généralisation des métriques de Wasserstein. Dans une seconde partie, nous développons des méthodes numériques pour résoudre les deux formulations et étudions en particulier une nouvelle famille d'algorithmes de "scaling", s'appliquant à une grande variété de problèmes. La troisième partie contient des illustrations ainsi que l'étude théorique et numérique, d'un flot de gradient de type Hele-Shaw dans l'espace des mesures. Pour les mesures à valeurs matricielles, nous proposons aussi un modèle de transport optimal qui permet un bon arbitrage entre fidélité géométrique et efficacité algorithmique. / This thesis generalizes optimal transport beyond the classical "balanced" setting of probability distributions. We define unbalanced optimal transport models between nonnegative measures, based either on the notion of interpolation or the notion of coupling of measures. We show relationships between these approaches. One of the outcomes of this framework is a generalization of the p-Wasserstein metrics. Secondly, we build numerical methods to solve interpolation and coupling-based models. We study, in particular, a new family of scaling algorithms that generalize Sinkhorn's algorithm. The third part deals with applications. It contains a theoretical and numerical study of a Hele-Shaw type gradient flow in the space of nonnegative measures. It also adresses the case of measures taking values in the cone of positive semi-definite matrices, for which we introduce a model that achieves a balance between geometrical accuracy and algorithmic efficiency.

Nous définissons des formes normales géodésiques pour les séries générales des groupes de réflexions complexes G(de,e,n). Ceci nécessite l'élaboration d'une technique combinatoire afin de déterminer des décompositions réduites et de calculer la longueur des éléments de G(de,e,n) sur un ensemble générateur donné. En utilisant ces formes normales géodésiques, nous construisons des intervalles dans G(e,e,n) qui permettent d'obtenir des groupes de Garside. Certains de ces groupes correspondent au groupe de tresses complexe B(e,e,n). Pour les autres groupes de Garside, nous étudions certaines de leurs propriétés et nous calculons leurs groupes d'homologie sur Z d'ordre 2. Inspirés par les formes normales géodésiques, nous définissons aussi de nouvelles présentations et de nouvelles bases pour les algèbres de Hecke associées aux groupes de réflexions complexes G(e,e,n) et G(d,1,n) ce qui permet d'obtenir une nouvelle preuve de la conjecture de liberté de BMR (Broué-Malle-Rouquier) pour ces deux cas. Ensuite, nous définissons des algèbres de BMW (Birman-Murakami-Wenzl) et de Brauer pour le type (e,e,n). Ceci nous permet de construire des représentations de Krammer explicites pour des cas particuliers des groupes de tresses complexes B(e,e,n). Nous conjecturons que ces représentations sont fidèles. Enfin, en se basant sur nos calculs heuristiques, nous proposons une conjecture sur la structure de l'algèbre de BMW. / We define geodesic normal forms for the general series of complex reflection groups G(de,e,n). This requires the elaboration of a combinatorial technique in order to determine minimal word representatives and to compute the length of the elements of G(de,e,n) over some generating set. Using these geodesic normal forms, we construct intervals in G(e,e,n) that give rise to Garside groups. Some of these groups correspond to the complex braid group B(e,e,n). For the other Garside groups that appear, we study some of their properties and compute their second integral homology groups. Inspired by the geodesic normal forms, we also define new presentations and new bases for the Hecke algebras associated to the complex reflection groups G(e,e,n) and G(d,1,n) which lead to a new proof of the BMR (Broué-Malle-Rouquier) freeness conjecture for these two cases. Next, we define a BMW (Birman-Murakami-Wenzl) and Brauer algebras for type (e,e,n). This enables us to construct explicit Krammer's representations for some cases of the complex braid groups B(e,e,n). We conjecture that these representations are faithful. Finally, based on our heuristic computations, we propose a conjecture about the structure of the BMW algebra.