Cette thèse s’intéresse à diverses composantes du traitement et de l’analyse d’images par méthodes non locales. Ces méthodes sont basées sur la redondance d’information présente dans d’autres images, et utilisent des algorithmes de recherche de correspondance, généralement basés sur l’utilisation patchs, pour extraire et transférer de l’information depuis ces images d’exemples. Ces approches, largement utilisées par la communauté de vision par ordinateur, sont souvent limitées par le temps de calcul de l’algorithme de recherche, appliqué à chaque pixel, et par la nécessité d’effectuer un prétraitement ou un apprentissage pour utiliser de grandes bases de données.Pour pallier ces limites, nous proposons plusieurs méthodes générales, sans apprentissage,rapides, et qui peuvent être facilement adaptées à diverses applications de traitement et d’analyse d’images naturelles ou médicales. Nous introduisons un algorithme de recherche de correspondances permettant d’extraire rapidement des patchs d’une grande bibliothèque d’images 3D, que nous appliquons à la segmentation d’images médicales. Pour utiliser de façon similaire aux patchs,des présegmentations en superpixels réduisant le nombre d’éléments de l’image,nous présentons une nouvelle structure de voisinage de superpixels. Ce nouveau descripteur permet d’utiliser efficacement les superpixels dans des approches non locales. Nous proposons également une méthode de décomposition régulière et précise en superpixels. Nous montrons comment évaluer cette régularité de façon robuste, et que celle-ci est nécessaire pour obtenir de bonnes performances de recherche de correspondances basées sur les superpixels. / This thesis focuses on several aspects of image analysis and processing with non local methods. These methods are based on the redundancy of information that occurs in other images, and use matching algorithms, that are usually patch-based, to extract and transfer information from the example data. These approaches are widely used by the computer vision community, and are generally limited by the computational time of the matching algorithm, applied at the pixel scale, and by the necessity to perform preprocessing or learning steps to use large databases. To address these issues, we propose several general methods, without learning, fast, and that can be easily applied to different image analysis and processing applications on natural and medical images. We introduce a matching algorithm that enables to quickly extract patches from a large library of 3D images, that we apply to medical image segmentation. To use a presegmentation into superpixels that reduces the number of image elements, in a way that is similar to patches, we present a new superpixel neighborhood structure. This novel descriptor enables to efficiently use superpixels in non local approaches. We also introduce an accurate and regular superpixel decomposition method. We show how to evaluate this regularity in a robust manner, and that this property is necessary to obtain good superpixel-based matching performances.

Uma das etapas mais importantes da análise de imagens e, que conta com uma enorme quantidade de aplicações, é a segmentação. No entanto, uma boa parte das técnicas tradicionais apresenta alto custo computacional, dificultando sua aplicação em imagens de alta resolução como, por exemplo, as imagens de ninhais de aves do Pantanal que também serão analisadas neste trabalho. Diante disso, é proposta uma nova abordagem de segmentação que combina algoritmos de detecção de comunidades, pertencentes à teoria das redes complexas, com técnicas de extração de superpixels. Tal abordagem é capaz de segmentar imagens de alta resolução mantendo o compromisso entre acurácia e tempo de processamento. Além disso, como as imagens de ninhais analisadas apresentam características peculiares que podem ser mais bem tratadas por técnicas de segmentação por textura, a técnica baseada em Markov Random Fields (MRF) é proposta, como um complemento à abordagem de segmentação inicial, para realizar a identificação final das aves. Por fim, devido à importância de avaliar quantitativamente a qualidade das segmentações obtidas, um nova métrica de avaliação baseada em ground-truth foi desenvolvida, sendo de grande importância para a área. Este trabalho contribuiu para o avanço do estado da arte das técnicas de segmentação de imagens de alta resolução, aprimorando e desenvolvendo métodos baseados na combinação de redes complexas com superpixels, os quais alcançaram resultados satisfatórios com baixo tempo de processamento. Além disso, uma importante contribuição referente ao censo demográfico de aves por meio da análise de imagens aéreas de ninhais foi viabilizada por meio da aplicação da técnica de segmentação MRF. / Segmentation is one of the most important steps in image analysis with a large range of applications. However, some traditional techniques exhibit high computational costs, hindering their application in high resolution images such as the images of birds nests from Pantanal, one of Brazilian most important wetlands. Therefore, we propose a new segmentation approach that combines community detection algorithms, originated from the theory of the complex networks, with superpixels extraction techniques. This approach is capable of segmenting high resolution images while maintaining the trade-off between accuracy and processing time. Moreover, as the nest images exhibit peculiar characteristics that can be better dealt with texture segmentation techniques, the Markov Random Fields (MRF) technique is proposed, as a complement to the initial approach, to perform the final identification of the birds. Finally, due to the importance of the quantitatively evaluation of the segmentation quality, a new evaluation metric based on ground-truth was developed, being of great importance to the segmentation field. This work contributed to the state of art of high resolution images segmentation techniques, improving and developing methods based on combination of complex networks and superpixels, which generated satisfactory results within low processing time. Moreover, an important contribution for the birds census by the analysis of aerial images of birds nests was made possible by application of the MRF technique.

Les entreprises doivent traiter quotidiennement de gros volumes de documents papiers de toutes sortes. Automatisation, traçabilité, alimentation de systèmes d’informations, réduction des coûts et des délais de traitement, la dématérialisation a un impact économique évident. Pour respecter les contraintes industrielles, les processus historiques d’analyse simplifient les images grâce à une séparation fond/premier-plan. Cependant, cette binarisation peut être source d’erreurs lors des étapes de segmentation et de reconnaissance. Avec l’amélioration des techniques, la communauté d’analyse de documents a montré un intérêt croissant pour l’intégration d’informations colorimétriques dans les traitements, ceci afin d’améliorer leurs performances. Pour respecter le cadre imposé par notre partenaire privé, l’objectif était de mettre en place des processus non supervisés. Notre but est d’être capable d’analyser des documents même rencontrés pour la première fois quels que soient leurs contenus, leurs structures, et leurs caractéristiques en termes de couleurs. Les problématiques de ces travaux ont été d’une part l’identification d’un nombre raisonnable de couleurs principales sur une image ; et d’autre part, le regroupement en couches couleur cohérentes des pixels ayant à la fois une apparence colorimétrique très proche, et présentant une unité logique ou sémantique. Fournies sous forme d’un ensemble d’images binaires, ces couches peuvent être réinjectées dans la chaîne de dématérialisation en fournissant une alternative à l’étape de binarisation classique. Elles apportent en plus des informations complémentaires qui peuvent être exploitées dans un but de segmentation, de localisation, ou de description. Pour cela, nous avons proposé une segmentation spatio-colorimétrique qui permet d’obtenir un ensemble de régions locales perceptuellement cohérentes appelées superpixels, et dont la taille s’adapte au contenu spécifique des images de documents. Ces régions sont ensuite regroupées en couches couleur globales grâce à une analyse multi-résolution. / Industrial companies receive huge volumes of documents everyday. Automation, traceability, feeding information systems, reducing costs and processing times, dematerialization has a clear economic impact. In order to respect the industrial constraints, the traditional digitization process simplifies the images by performing a background/foreground separation. However, this binarization can lead to some segmentation and recognition errors. With the improvements of technology, the community of document analysis has shown a growing interest in the integration of color information in the process to enhance its performance. In order to work within the scope provided by our industrial partner in the digitization flow, an unsupervised segmentation approach was chosen. Our goal is to be able to cope with document images, even when they are encountered for the first time, regardless their content, their structure, and their color properties. To this end, the first issue of this project was to identify a reasonable number of main colors which are observable on an image. Then, we aim to group pixels having both close color properties and a logical or semantic unit into consistent color layers. Thus, provided as a set of binary images, these layers may be reinjected into the digitization chain as an alternative to the conventional binarization. Moreover, they also provide extra-information about colors which could be exploited for segmentation purpose, elements spotting, or as a descriptor. Therefore, we have proposed a spatio-colorimetric approach which gives a set of local regions, known as superpixels, which are perceptually meaningful. Their size is adapted to the content of the document images. These regions are then merged into global color layers by means of a multiresolution analysis.

Le rendu photo-réaliste traditionnel exige un effort manuel et des calculs intensifs pour créer des scènes et rendre des images réalistes. C'est principalement pour cette raison que la création de contenus pour l’imagerie numérique de haute qualité a été limitée aux experts et le rendu hautement réaliste nécessite encore des temps de calcul significatifs. Le rendu basé image (IBR) est une alternative qui a le potentiel de rendre les applications de création et de rendu de contenus de haute qualité accessibles aux utilisateurs occasionnels, puisqu'ils peuvent générer des images photo-réalistes de haute qualité sans subir les limitations mentionnées ci-dessus. Nous avons identifié trois limitations importantes des méthodes actuelles de rendu basé image : premièrement, chaque algorithme possède des forces et faiblesses différentes, en fonction de la qualité de la reconstruction 3D et du contenu de la scène, et un seul algorithme ne permet souvent pas d’obtenir la meilleure qualité de rendu partout dans l’image. Deuxièmement, ces algorithmes présentent de forts artefacts lors du rendu d’objets manquants ou partiellement reconstruits. Troisièmement, la plupart des méthodes souffrent encore d'artefacts visuels significatifs dans les régions de l’image où la reconstruction est de faible qualité. Dans l'ensemble, cette thèse propose plusieurs améliorations significatives du rendu basé image aussi bien en termes de vitesse de rendu que de qualité d’image. Ces nouvelles solutions sont basées sur le rendu sélectif, la substitution de modèle basé sur l'apprentissage, et la prédiction et la correction des erreurs de profondeur. / Traditional photo-realistic rendering requires intensive manual and computational effort to create scenes and render realistic images. Thus, creation of content for high quality digital imagery has been limited to experts and highly realistic rendering still requires significant computational time. Image-Based Rendering (IBR) is an alternative which has the potential of making high-quality content creation and rendering applications accessible to casual users, since they can generate high quality photo-realistic imagery without the limitations mentioned above. We identified three important shortcomings of current IBR methods: First, each algorithm has different strengths and weaknesses, depending on 3D reconstruction quality and scene content and often no single algorithm offers the best image quality everywhere in the image. Second, such algorithms present strong artifacts when rendering partially reconstructed objects or missing objects. Third, most methods still result in significant visual artifacts in image regions where reconstruction is poor. Overall, this thesis addresses significant shortcomings of IBR for both speed and image quality, offering novel and effective solutions based on selective rendering, learning-based model substitution and depth error prediction and correction.

Uma das etapas mais importantes da análise de imagens e, que conta com uma enorme quantidade de aplicações, é a segmentação. No entanto, uma boa parte das técnicas tradicionais apresenta alto custo computacional, dificultando sua aplicação em imagens de alta resolução como, por exemplo, as imagens de ninhais de aves do Pantanal que também serão analisadas neste trabalho. Diante disso, é proposta uma nova abordagem de segmentação que combina algoritmos de detecção de comunidades, pertencentes à teoria das redes complexas, com técnicas de extração de superpixels. Tal abordagem é capaz de segmentar imagens de alta resolução mantendo o compromisso entre acurácia e tempo de processamento. Além disso, como as imagens de ninhais analisadas apresentam características peculiares que podem ser mais bem tratadas por técnicas de segmentação por textura, a técnica baseada em Markov Random Fields (MRF) é proposta, como um complemento à abordagem de segmentação inicial, para realizar a identificação final das aves. Por fim, devido à importância de avaliar quantitativamente a qualidade das segmentações obtidas, um nova métrica de avaliação baseada em ground-truth foi desenvolvida, sendo de grande importância para a área. Este trabalho contribuiu para o avanço do estado da arte das técnicas de segmentação de imagens de alta resolução, aprimorando e desenvolvendo métodos baseados na combinação de redes complexas com superpixels, os quais alcançaram resultados satisfatórios com baixo tempo de processamento. Além disso, uma importante contribuição referente ao censo demográfico de aves por meio da análise de imagens aéreas de ninhais foi viabilizada por meio da aplicação da técnica de segmentação MRF. / Segmentation is one of the most important steps in image analysis with a large range of applications. However, some traditional techniques exhibit high computational costs, hindering their application in high resolution images such as the images of birds nests from Pantanal, one of Brazilian most important wetlands. Therefore, we propose a new segmentation approach that combines community detection algorithms, originated from the theory of the complex networks, with superpixels extraction techniques. This approach is capable of segmenting high resolution images while maintaining the trade-off between accuracy and processing time. Moreover, as the nest images exhibit peculiar characteristics that can be better dealt with texture segmentation techniques, the Markov Random Fields (MRF) technique is proposed, as a complement to the initial approach, to perform the final identification of the birds. Finally, due to the importance of the quantitatively evaluation of the segmentation quality, a new evaluation metric based on ground-truth was developed, being of great importance to the segmentation field. This work contributed to the state of art of high resolution images segmentation techniques, improving and developing methods based on combination of complex networks and superpixels, which generated satisfactory results within low processing time. Moreover, an important contribution for the birds census by the analysis of aerial images of birds nests was made possible by application of the MRF technique.

Le cancer de la prostate est une cause majeure de décès dans le monde. La radiothérapie externe est une des techniques utilisée pour traiter ce cancer. Pour ce faire, la segmentation de la prostate et de ses organes à risque (OAR) associés (le rectum, la vessie et les têtes fémorales) est une étape majeure dans l’application du traitement. L’objectif de cette thèse est de fournir des outils afin de segmenter la prostate et les OAR de manière automatique ou semi-automatique. Plusieurs approches ont été proposées ces dernières années pour répondre à ces problématiques. Les OAR possédant un contraste relativement bon dans l’image, nous nous sommes orientés vers une approche semi-automatique de leur segmentation, consistant en une sur-segmentation de l’image en petites régions homogènes appelées superpixels. L’utilisateur de la méthode choisit ensuite de labelliser quelques superpixels dans les OAR comme des germes. Enfin, la méthode segmente les OAR grâce à une diffusion sur le graphe (à partir des germes) construit par des superpixels. Quant à la segmentation de la prostate, un sous-volume de l’image appelé VOI (Volume Of Interest), dans lequel se trouve la prostate, est tout d’abord défini. À l’intérieur de ce VOI, la segmentation de la prostate est réalisée. Un dictionnaire composé des caractéristiques de textures extraites sur chaque patch du VOI est d’abord construit. La sélection de caractéristiques du dictionnaire sous contraintes parcimonieuses permet ensuite de trouver celles qui sont le plus informatives. Enfin, basé sur ces caractéristiques sélectionnées, une propagation de label de patch sous contrainte parcimonieuse est appliquée pour segmenter la prostate à deux échelles, superpixels et pixels. Notre méthode a été évaluée sur des images TDM du Centre Henri Becquerel et IRM du challenge ISBI 2013 avec des résultats prometteurs. / Prostate cancer is a leading cause of death worldwide. External radiotherapy is one of the techniques used to this disease. In order to achieve this, the segmentation of the prostate and its associated organs at risk (OAR) (rectum, bladder and femoral heads) is a major step in the application of the treatment. The objective of this thesis is to provide tools to segment prostate and OAR automatically or semi-automatically. Several approaches have been proposed in recent years to address these issues. As OAR have a relatively good contrast in the image, we have focused on a semi-automatic approach to segment them, consisting of an over-segmentation of the image into small homogeneous regions called superpixels. Then, the user labels some superpixels in the OAR as germs. Finally, the OAR segmentation is performed by a graph diffusion (from germs) constructed by superpixels. Regarding the prostate segmentation, a sub-volume of the image called VOI (Volume Of Interest), in which the prostate is located, is first defined. The prostate segmentation is performed within this VOI. A dictionary composed of the texture characteristics extracted on each patch of the VOI is first constructed. Then, the selection of characteristics of the dictionary under parsimonious constraints allows to find the most informative ones. Finally, based on these selected characteristics, patch label propagation under parsimonious constraint is applied to segment the prostate at two scales, superpixels and pixels. Our method was evaluated with promising results on TDM images of the Henri Becquerel Center and IRM of the 2013 ISBI challenge.

Immunohistochemical (IHC) assessment in cancer research is important for understanding the distribution and localisation of biomarkers at the cellular level. However currently IHC analyses are predominantly performed manually, increasing workloads and introducing inter- and intra-observer variability. Automation shows great potential in clinical research to reduce pathologists' workloads and speed up cancer research in large clinical studies. Whilst recent advancements in digital pathology have enabled IHC measurements to be performed automatically, the acquisition of manual annotations of tumours in scanned digital slides is still a limiting factor. In this thesis, an automated solution to tumour localisation is explored with the aim of replacing manual annotations. As an exemplar, human breast tissue microarrays stained with estrogen receptor are considered. Methods for automated tumour localisation are described with a focus on capturing structural information in tissue by adopting superpixel properties in a rotation invariant manner, suitable for histopathology images. To incorporate essential contextual information, methods which utilise posterior tumour probabilities in an iterative manner are proposed. Results showed pixel-level agreements between automated and manual tumour segmentation masks (κ=0.811) approach inter-rater agreement between expert pathologists (κ=0.908). A large proportion of disagreements between automated and manual segmentations were shown to correlate to minor discrepancies, inconsequential for IHC assessment. IHC scores extracted from automated and manual tumour segmentation masks showed strong agreements (Allred: κˆ=0.911; Quickscore: κˆ=0.922), demonstrating the potential of automation in clinical practice across large clinical trials.

Over the last few years, the concept of civil Unmanned Aircraft System(s) (UAS) has been realised, with small UASs commonly used in industries such as law enforcement, agriculture and mapping. With increased development in other areas, such as logistics and advertisement, the size and range of civil UAS is likely to grow. Taken to the logical conclusion, it is likely that large scale UAS will be operating in civil airspace within the next decade. Although the airborne operations of civil UAS have already gathered much research attention, work is also required to determine how UAS will function when on the ground. Motivated by the assumption that large UAS will share ground facilities with manned aircraft, this thesis describes the preliminary development of an Automated Taxiing System(ATS) for UAS operating at civil aerodromes. To allow the ATS to function on the majority of UAS without the need for additional hardware, a visual sensing approach has been chosen, with the majority of work focusing on monocular image processing techniques. The purpose of the computer vision system is to provide direct sensor data which can be used to validate the vehicle s position, in addition to detecting potential collision risks. As aerospace regulations require the most robust and reliable algorithms for control, any methods which are not fully definable or explainable will not be suitable for real-world use. Therefore, non-deterministic methods and algorithms with hidden components (such as Artificial Neural Network (ANN)) have not been used. Instead, the visual sensing is achieved through a semantic segmentation, with separate segmentation and classification stages. Segmentation is performed using superpixels and reachability clustering to divide the image into single content clusters. Each cluster is then classified using multiple types of image data, probabilistically fused within a Bayesian network. The data set for testing has been provided by BAE Systems, allowing the system to be trained and tested on real-world aerodrome data. The system has demonstrated good performance on this limited dataset, accurately detecting both collision risks and terrain features for use in navigation.

[pt] A segmentação de imagens com objetivo de determinar a forma de objetos é ainda um problema difícil. A separação de regiões que correspondem a objetos contidos na imagem geralmente leva em consideração propriedades de similaridade, proximidade e descontinuidade. A imagem a ser segmentada pode ser de diversas naturezas, como fotografias, imagens médicas e sísmicas. Podemos encontrar na literatura muitos métodos de segmentação propostos como possíveis soluções para diferentes problemas. Recentemente a técnica de superpixel tem sido utilizada como um passo inicial que reduz o tamanho da entrada do problema. Este trabalho propõe uma metodologia de segmentação de imagens fotográficas e de ultrassom que se baseia em variantes de superpixels. A metodologia proposta se adapta a natureza da imagem e a complexidade do problema utilizando diferentes medidas de similaridade e distância. O trabalho apresenta também resultados que buscam esclarecer o procedimento proposto e a escolha de seus parâmetros. / [en] Image segmentation for object modeling is a complex task that is still not well solved. The separation of the regions corresponding to each object in an image is based on proximity, similarity, and discontinuity of its boundaries. The image to be segmented can be of various natures, including photographs, medical and seismic images. We can find in literature many proposed segmentation methods used as solutions to different problems. Recently the superpixel technique has been used as an initial step that reduces the size of the problem input. This work proposes a methodology of segmentation of photographs and ultrasound images based on variants of superpixels. The proposed methodology adapts to the image s nature and to the problem s complexity using different measures of similarity and distance. This work also presents results that seek to clarify the proposed procedure and the choice of its parameters.

Title: Gradient boosted segmentation of retinal fundus images Author: Matúš Goliaš Department: Department of Software and Computer Science Education Supervisor: Doc. RNDr. Elena Šikudová PhD., Department of Software and Computer Science Education Abstract: Over the recent years, there has been an increase in the use of automatic methods in medical diagnosis. A significant number of publications have analysed eye disorders and diseases. One of the most severe eye conditions is glaucoma. It damages optic nerves and causes gradual loss of vision. An essential step towards a faster diagnosis of this disease is accurate segmentation of the optic disc and cup. This task is difficult due to many retinal defects, different image acquisition techniques, and artefacts caused by imaging devices. This thesis describes an iterative threshold-based algorithm for extraction of the optic disc. An objective function quantifying object similarity to the optic disc is defined to direct the iteration. Following that, we introduce a superpixel-based classification algorithm for extraction of the optic cup. We propose the use of gradient boosted decision trees which outperform random forest and support vector machine. In addition, we evaluate the proposed algorithms and their alternatives on a publicly available retinal fundus...