Dans cette thèse, nous nous intéressons à la recherche d’images similaires avec des descripteurs structurés par découpages d’images sur les graphes.Nous proposons une nouvelle approche appelée “bag-of-bags of words” (BBoW) pour la recherche d’images par le contenu (CBIR). Il s’agit d’une extension du modèle classique dit sac-de-mots (bag of words - BoW). Dans notre approche, une image est représentée par un graphe placé sur une grille régulière de pixels d’image. Les poids sur les arêtes dépendent de caractéristiques locales de couleur et texture. Le graphe est découpé en un nombre fixe de régions qui constituent une partition irrégulière de l’image. Enfin, chaque partition est représentée par sa propre signature suivant le même schéma que le BoW. Une image est donc décrite par un ensemble de signatures qui sont ensuite combinées pour la recherche d’images similaires dans une base de données. Contrairement aux méthodes existantes telles que Spatial Pyramid Matching (SPM), le modèle BBoW proposé ne repose pas sur l’hypothèse que des parties similaires d’une scène apparaissent toujours au même endroit dans des images d’une même catégorie. L’extension de cette méthode ` a une approche multi-échelle, appelée Irregular Pyramid Matching (IPM), est ´ également décrite. Les résultats montrent la qualité de notre approche lorsque les partitions obtenues sont stables au sein d’une même catégorie d’images. Une analyse statistique est menée pour définir concrètement la notion de partition stable.Nous donnons nos résultats sur des bases de données pour la reconnaissance d’objets, d’indexation et de recherche d’images par le contenu afin de montrer le caractère général de nos contributions / Image representation is a fundamental question for several computer vision tasks. The contributions discussed in this thesis extend the basic bag-of-words representations for the tasks of object recognition and image retrieval.In the present thesis, we are interested in image description by structural graph descriptors. We propose a model, named bag-of-bags of words (BBoW), to address the problems of object recognition (for object search by similarity), and especially Content-Based Image Retrieval (CBIR) from image databases. The proposed BBoW model, is an approach based on irregular pyramid partitions over the image. An image is first represented as a connected graph of local features on a regular grid of pixels. Irregular partitions (subgraphs) of the image are further built by using graph partitioning methods. Each subgraph in the partition is then represented by its own signature. The BBoW model with the aid of graphs, extends the classical bag-of-words (BoW) model by embedding color homogeneity and limited spatial information through irregular partitions of an image. Compared to existing methods for image retrieval, such as Spatial Pyramid Matching (SPM), the BBoW model does not assume that similar parts of a scene always appear at the same location in images of the same category. The extension of the proposed model to pyramid gives rise to a method we named irregular pyramid matching (IPM).The experiments demonstrate the strength of our approach for image retrieval when the partitions are stable across an image category. The statistical analysisof subgraphs is fulfilled in the thesis. To validate our contributions, we report results on three related computer vision datasets for object recognition, (localized)content-based image retrieval and image indexing. The experimental results in a database of 13,044 general-purposed images demonstrate the efficiency and effectiveness of the proposed BBoW framework.

Many different approaches have been taken towards solving the stereo correspondence problem and great progress has been made within the field during the last decade. This is mainly thanks to newly evolved global optimization techniques and better ways to compute pixel dissimilarity between views. The most successful algorithms are based on approaches that explicitly model smoothness assumptions made about the physical world, with image segmentation and plane fitting being two frequently used techniques. Within the project, a survey of state of the art stereo algorithms was conducted and the theory behind them is explained. Techniques found interesting were implemented for experimental trials and an algorithm aiming to achieve state of the art performance was implemented and evaluated. For several cases, state of the art performance was reached. To keep down the computational complexity, an algorithm relying on local winner-take-all optimization, image segmentation and plane fitting was compared against minimizing a global energy function formulated on pixel level. Experiments show that the local approach in several cases can match the global approach, but that problems sometimes arise – especially when large areas that lack texture are present. Such problematic areas are better handled by the explicit modeling of smoothness in global energy minimization. Lastly, disparity estimation for image sequences was explored and some ideas on how to use temporal information were implemented and tried. The ideas mainly relied on motion detection to determine parts that are static in a sequence of frames. Stereo correspondence for sequences is a rather new research field, and there is still a lot of work to be made.

Nous développons une méthode générique semi-automatique multi-objet de segmentation d'image par coupure de graphe visant les usages médicaux de routine, allant des tâches impliquant quelques objets dans des images 2D, à quelques dizaines dans celles 3D quasi corps entier. La formulation souple de la méthode permet son adaptation simple à une application donnée. En particulier, le modèle d'a priori de proximité que nous proposons, défini à partir des contraintes de paires du plus court chemin sur le graphe d'adjacence des objets, peut facilement être adapté pour tenir compte des relations spatiales entre les objets ciblés dans un problème donné. L'algorithme de segmentation peut être adapté aux besoins de l'application en termes de temps d'exécution et de capacité de stockage à l'aide d'une partition de l'image à segmenter par une tesselation de Voronoï efficace et contrôlable, établissant un bon équilibre entre la compacité des régions et le respect des frontières des objets. Des évaluations et comparaisons qualitatives et quantitatives avec le modèle de Potts standard confirment que notre modèle d'a priori apporte des améliorations significatives dans la segmentation d'objets distincts d'intensités similaires, dans le positionnement précis des frontières des objets ainsi que dans la robustesse de segmentation par rapport à la résolution de partition. L'évaluation comparative de la méthode de partition avec ses concurrentes confirme ses avantages en termes de temps d'exécution et de qualité des partitions produites. Par comparaison avec l'approche appliquée directement sur les voxels de l'image, l'étape de partition améliore à la fois le temps d'exécution global et l'empreinte mémoire du processus de segmentation jusqu'à un ordre de grandeur, sans compromettre la qualité de la segmentation en pratique. / We develop a generic Graph Cut-based semiautomatic multiobject image segmentation method principally for use in routine medical applications ranging from tasks involving few objects in 2D images to fairly complex near whole-body 3D image segmentation. The flexible formulation of the method allows its straightforward adaption to a given application.\linebreak In particular, the graph-based vicinity prior model we propose, defined as shortest-path pairwise constraints on the object adjacency graph, can be easily reformulated to account for the spatial relationships between objects in a given problem instance. The segmentation algorithm can be tailored to the runtime requirements of the application and the online storage capacities of the computing platform by an efficient and controllable Voronoi tessellation clustering of the input image which achieves a good balance between cluster compactness and boundary adherence criteria. Qualitative and quantitative comprehensive evaluation and comparison with the standard Potts model confirm that the vicinity prior model brings significant improvements in the correct segmentation of distinct objects of identical intensity, the accurate placement of object boundaries and the robustness of segmentation with respect to clustering resolution. Comparative evaluation of the clustering method with competing ones confirms its benefits in terms of runtime and quality of produced partitions. Importantly, compared to voxel segmentation, the clustering step improves both overall runtime and memory footprint of the segmentation process up to an order of magnitude virtually without compromising the segmentation quality.

Magnetic resonance images (MRI) are acquired by sampling the current of induced electromotiveforce (EMF). EMF is induced due to flux of the net magnetic field from coherent nuclear spins with intrinsic magnetic dipole moments. The spins are excited by (non-ionizing) radio frequency electromagnetic radiation in conjunction with stationary and gradient magnetic fields. These images reveal detailed internal morphological structures as well as enable functional assessment of the body that can help diagnose a wide range of medical conditions. The aim of this project was to unwrap phase contrast cine magnetic resonance images, targeting the great vessels. The maximum encoded velocity (venc) is limited to the angular phase range [-π, π] radians. This may result in aliasing if the venc is set too low by the MRI personnel. Aliased images yield inaccurate cardiac stroke volume measurements and therefore require acquisition retakes. The retakes might be avoided if the images could be unwrapped in post-processing instead. Using computer vision, the angular phase of flow measurements as well as the angular phase of retrospectively wrapped image sets were unwrapped. The performances of three algorithms were assessed, Laplacian algorithm, sequential tree-reweighted message passing and iterative graph cuts. The associated energy formulation was also evaluated. Iterative graph cuts was shown to be the most robust with respect to the number of wraps and the energies correlated with the errors. This thesis shows that there is potential to reduce the number of acquisition retakes, although the MRI personnel still need to verify that the unwrapping performances are satisfactory. Given the promising results of iterative graph cuts, next it would be valuable to investigate the performance of a globally optimal surface estimation algorithm.