The main idea of this thesis is to present the novel concept of localized feature selection (LFS) for data classification and its application for coma outcome prediction. Typical feature selection methods choose an optimal global feature subset that is applied over all regions of the sample space. In contrast, in this study we propose a novel localized feature selection approach whereby each region of the sample space is associated with its own distinct optimized feature set, which may vary both in membership and size across the sample space. This allows the feature set to optimally adapt to local variations in the sample space. An associated localized classification method is also proposed. The proposed LFS method selects a feature subset such that, within a localized region, within-class and between-class distances are respectively minimized and maximized. We first determine the localized region using an iterative procedure based on the distances in the original feature space. This results in a linear programming optimization problem. Then, the second method is formulated as a non-linear joint convex/increasing quasi-convex optimization problem where a logistic function is applied to focus the optimization process on the localized region within the unknown co-ordinate system. This results in a more accurate classification performance at the expense of some sacrifice in computational time. Experimental results on synthetic and real-world data sets demonstrate the effectiveness of the proposed localized approach. Using the LFS idea, we propose a practical machine learning approach for automatic and continuous assessment of event related potentials for detecting the presence of the mismatch negativity component, whose existence has a high correlation with coma awakening. This process enables us to determine prognosis of a coma patient. Experimental results on normal and comatose subjects demonstrate the effectiveness of the proposed method. / Dissertation / Doctor of Philosophy (PhD) / This study proposes a novel form of pattern classification method, which is formulated in a way so that it is easily executable on a computer. Two different versions of the method are developed. These are the LFS (localized feature selection) and lLFS (logistic LFS) methods. Both versions are appropriate for analysis of data with complex distributions, such as datasets that occur in biological signal processing problems. We have shown that the performance of the proposed methods is significantly improved over that of previous methods, on the datasets that were considered in this thesis. The proposed method is applied to the specific problem of determining the prognosis of a coma patient. The viability of the formulation and the effectiveness of the proposed algorithm are demonstrated on several synthetic and real world datasets, including comatose subjects.

Local feature detection and its applications in different problems are quite popular in vision research. In order to analyze a scene, its invariant features, which are distinguishable in many views of this scene, are used in pose estimation, object detection and augmented reality. However, required performance metrics might change according to the application type / in general, the main metrics are accepted as accuracy and computational complexity. The contributions in this thesis provide improving these metrics and can be divided into three parts, as local feature detection, local feature description and description matching in different views of the same scene. In this thesis an efficient feature detection algorithm with sufficient repeatability performance is proposed. This detection method is convenient for real-time applications. For local description, a novel local binary pattern outperforming state-of-the-art binary pattern is proposed. As a final task, a fuzzy decision tree method is presented for approximate nearest neighbor search. In all parts of the system, computational efficiency is considered and the algorithms are designed according to limited processing time. Finally, an overall system capable of matching different views of the same scene has been proposed and executed in a mobile platform. The results are quite promising such that the presented system can be used in real-time applications, such as augmented reality, object retrieval, object tracking and pose estimation.

Augmented Reality (AR) involves adding virtual content into real scenes. Scenes are viewed using a Head-Mounted Display or other display type. In order to place content into the user's view of a scene, the user's position and orientation relative to the scene, commonly referred to as their pose, must be determined accurately. This allows the objects to be placed in the correct positions and to remain there when the user moves or the scene changes. It is achieved by tracking the user in relation to their environment using a variety of technology. One technology which has proven to provide accurate results is computer vision. Computer vision involves a computer analysing images and achieving an understanding of them. This may be locating objects such as faces in the images, or in the case of AR, determining the pose of the user. One of the ultimate goals of AR systems is to be capable of operating under any condition. For example, a computer vision system must be robust under a range of different scene types, and under unpredictable environmental conditions due to variable illumination and weather. The majority of existing literature tests algorithms under the assumption of ideal or 'normal' imaging conditions. To ensure robustness under as many circumstances as possible it is also important to evaluate the systems under adverse conditions. This thesis seeks to analyse the effects that variable illumination has on computer vision algorithms. To enable this analysis, test data is required to isolate weather and illumination effects, without other factors such as changes in viewpoint that would bias the results. A new dataset is presented which also allows controlled viewpoint differences in the presence of weather and illumination changes. This is achieved by capturing video from a camera undergoing a repeatable motion sequence. Ground truth data is stored per frame allowing images from the same position under differing environmental conditions, to be easily extracted from the videos. An in depth analysis of six detection algorithms and five matching techniques demonstrates the impact that non-uniform illumination changes can have on vision algorithms. Specifically, shadows can degrade performance and reduce confidence in the system, decrease reliability, or even completely prevent successful operation. An investigation into approaches to improve performance yields techniques that can help reduce the impact of shadows. A novel algorithm is presented that merges reference data captured at different times, resulting in reference data with minimal shadow effects. This can significantly improve performance and reliability when operating on images containing shadow effects. These advances improve the robustness of computer vision systems and extend the range of conditions in which they can operate. This can increase the usefulness of the algorithms and the AR systems that employ them.

隨著智慧型手機的普及，在移動裝置上直接處理圖像的需求也大幅增加，故對於影像特徵描述子的要求，除了要表現出區域特徵的穩健性，同時也要維持良好的特徵比對效率與合理的儲存空間。過去所提出的區域影像特徵描述子建構方法之中，LIOP方法具有相當不錯的表現力，但其特徵描述子維度會隨著點群取樣數量的提高而以倍數增加，因此本研究提出Dynamic Local Intensity Order Relations (DLIOR)特徵描述子建構方法，利用LIOR方法探討點群中點與點之間的關係，減緩其維度增長幅度；透過動態設定像素差距門檻值，處理影像間像素差距分佈不均的問題，並使用線性轉換、點對歐幾里德距離等方式，重新定義描述子欄位的權重設定。經過實驗證實，DLIOR方法能夠使用比LIOP方法更少的維度空間，描述更多點群數的特徵資訊，並且具有更高的特徵比對能力。 / With the popularity of smart phones, the amounts of images being captured and processed on mobile devices have grown significantly in recent years. Image feature descriptors, which play crucial roles in recognition tasks, are expected to exhibit robust matching performance while at the same time maintain reasonable storage requirement. Among the local feature descriptors that have been proposed previously, local intensity order patterns (LIOP) demonstrated superior performance in many benchmark studies. As LIOP encodes the ranking relation in a point set (with N elements), however, its feature dimension increases drastically (N!) with the number of the neighboring sampling points around a pixel. To alleviate the dimensionality issue, this thesis presents a local feature descriptor by considering pairwise intensity relation in a pixel group, thereby reducing feature dimension to the order of C^N_2. In the proposed method, the threshold for assigning order relation is set dynamically according to local intensity distribution. Different weighting schemes, including linear transformation and Euclidean distance, have also been investigated to adjust the contribution of each pairing relation. Ultimately, the dynamic local intensity order relations (DLIOR) is devised to effectively encode intensity order relation of each pixel group. Experimental results indicate that DLIOR consumes less storage space than LIOP but achieves better feature matching performance using benchmark dataset.

La représentation visuelle est un problème fondamental en vision par ordinateur. Le but est de réduire l'information au strict nécessaire pour une tâche désirée. Plusieurs types de représentation existent, comme les caractéristiques de couleur (histogrammes, attributs de couleurs...), de forme (dérivées, points d'intérêt...) ou d'autres, comme les bancs de filtres.Les caractéristiques bas-niveau (locales) sont rapides à calculer. Elles ont un pouvoir de représentation limité, mais leur généricité présente un intérêt pour des systèmes autonomes et multi-tâches, puisque les caractéristiques haut-niveau découlent d'elles.Le but de cette thèse est de construire puis d'étudier l'impact de représentations fondées seulement sur des caractéristiques locales de bas-niveau (couleurs, dérivées spatiales) pour deux tâches : la poursuite d'objets génériques, nécessitant des caractéristiques robustes aux variations d'aspect de l'objet et du contexte au cours du temps; la détection d'objets, où la représentation doit décrire une classe d'objets en tenant compte des variations intra-classe. Plutôt que de construire des descripteurs d'objets globaux dédiés, nous nous appuyons entièrement sur les caractéristiques locales et sur des mécanismes statistiques flexibles visant à estimer leur distribution (histogrammes) et leurs co-occurrences (Transformée de Hough Généralisée). La Transformée de Hough Généralisée (THG), créée pour la détection de formes quelconques, consiste à créer une structure de données représentant un objet, une classe... Cette structure, d'abord indexée par l'orientation du gradient, a été étendue à d'autres caractéristiques. Travaillant sur des caractéristiques locales, nous voulons rester proche de la THG originale.En poursuite d'objets, après avoir présenté nos premiers travaux, combinant la THG avec un filtre particulaire (utilisant un histogramme de couleurs), nous présentons un algorithme plus léger et rapide (100fps), plus précis et robuste. Nous présentons une évaluation qualitative et étudierons l'impact des caractéristiques utilisées (espace de couleur, formulation des dérivées partielles...). En détection, nous avons utilisé l'algorithme de Gall appelé forêts de Hough. Notre but est de réduire l'espace de caractéristiques utilisé par Gall, en supprimant celles de type HOG, pour ne garder que les dérivées partielles et les caractéristiques de couleur. Pour compenser cette réduction, nous avons amélioré deux étapes de l'entraînement : le support des descripteurs locaux (patchs) est partiellement produit selon une mesure géométrique, et l'entraînement des nœuds se fait en générant une carte de probabilité spécifique prenant en compte les patchs utilisés pour cette étape. Avec l'espace de caractéristiques réduit, le détecteur n'est pas plus précis. Avec les mêmes caractéristiques que Gall, sur une même durée d'entraînement, nos travaux ont permis d'avoir des résultats identiques, mais avec une variance plus faible et donc une meilleure répétabilité. / Visual representation is a fundamental problem in computer vision. The aim is to reduce the information to the strict necessary for a query task. Many types of representation exist, like color features (histograms, color attributes...), shape ones (derivatives, keypoints...) or filterbanks.Low-level (and local) features are fast to compute. Their power of representation are limited, but their genericity have an interest for autonomous or multi-task systems, as higher level ones derivate from them. We aim to build, then study impact of low-level and local feature spaces (color and derivatives only) for two tasks: generic object tracking, requiring features robust to object and environment's aspect changes over the time; object detection, for which the representation should describe object class and cope with intra-class variations.Then, rather than using global object descriptors, we use entirely local features and statisticals mecanisms to estimate their distribution (histograms) and their co-occurrences (Generalized Hough Transform).The Generalized Hough Transform (GHT), created for detection of any shape, consists in building a codebook, originally indexed by gradient orientation, then to diverse features, modeling an object, a class. As we work on local features, we aim to remain close to the original GHT.In tracking, after presenting preliminary works combining the GHT with a particle filter (using color histograms), we present a lighter and fast (100 fps) tracker, more accurate and robust.We present a qualitative evaluation and study the impact of used features (color space, spatial derivative formulation).In detection, we used Gall's Hough Forest. We aim to reduce Gall's feature space and discard HOG features, to keep only derivatives and color ones.To compensate the reduction, we enhanced two steps: the support of local descriptors (patches) are partially chosen using a geometrical measure, and node training is done by using a specific probability map based on patches used at this step.With reduced feature space, the detector is less accurate than with Gall's feature space, but for the same training time, our works lead to identical results, but with higher stability and then better repeatability.

L'Accident Vasculaire Cérébral (AVC) - pathologie résultant d'une perturbation de l'apport sanguin dans le cerveau - est un problème de santé publique majeur, représentant la troisième cause de mortalité dans les pays industrialisés. Afin d'améliorer la prise en charge des patients atteints d'un AVC, il est important de posséder des méthodes efficaces pour l'identification des patients éligibles aux différentes thérapies et pour l'évaluation du rapport bénéfice/risque associé à ces thérapies. Dans ce contexte, l'Imagerie par Résonance Magnétique (IRM) dynamique de perfusion par contraste de susceptibilité, une modalité d'imagerie utile pour apprécier l'état de la perfusion cérébrale, peut aider à identifier les tissus à risque de s'infarcir. Cependant, l'intégralité de la chaîne de traitement, de l'acquisition à l'analyse et l'interprétation de l'IRM de perfusion demeure complexe et plusieurs limitations restent encore à surmonter. Durant ces travaux de thèse, nous contribuons à l'amélioration de la chaîne de traitement de l'IRM de perfusion, avec comme objectif final, l'obtention d'une meilleure prédiction de l'évolution de la lésion ischémique dans l'AVC. Dans une première partie, nous travaillons principalement sur l'étape de déconvolution des signaux temporels, une des étapes clefs à l'amélioration de l'IRM de perfusion. Cette étape consiste en la résolution d'un problème inverse mal-posé, et permet le calcul de paramètres hémodynamiques qui sont des biomarqueurs importants pour la classification de l'état final des tissus dans l'AVC. Afin de comparer de façon objective les performances des différents algorithmes de déconvolution existants et d'en valider des nouveaux, il est nécessaire d'avoir accès à une information sur la vérité terrain après déconvolution. Dans ce but, nous avons développé un simulateur numérique pour l'IRM de perfusion, avec une vérité terrain générée automatiquement. Ce simulateur est utilisé pour démontrer la faisabilité d'une automatisation du réglage des paramètres de régularisation, et établir la robustesse d'un algorithme de déconvolution avec régularisation spatio-temporelle d'introduction récente. Nous proposons également un nouvel algorithme de déconvolution globalement convergent. Enfin, la première partie de ces travaux se termine avec une discussion sur une autre étape de la chaîne de traitement en IRM de perfusion, à savoir, la normalisation des cartes de paramètres hémodynamiques extraites des images déconvoluées / Stroke – a neurological deficit resulting from blood supply perturbations in the brain – is a major public health issue, representing the third cause of death in industrialized countries. There is a need to improve the identification of patients eligible to the different therapies, as well as the evaluation of the benefit-risk ratio for the patients. In this context, perfusion Dynamic Susceptibility Contrast (DSC)-MRI, a prominent imaging modality for the assessment of cerebral perfusion, can help to identify the tissues at risk of infarction from the benign oligaemia. However, the entire pipeline from the acquisition to the analysis and interpretation of a DSC-MRI remains complex and some limitations are still to be overcome. During this PhD work, we contribute to improving the DSC-MRI processing pipeline with the ultimate objective of ameliorating the prediction of the ischemic lesion evolution in stroke. In a first part, we primarily work on the step of temporal signal deconvolution, one of the steps key to the improvement of DSC-MRI. This step consists in the resolution of an inverse ill-posed problem and allows the computation of hemodynamic parameters which are important biomarkers for tissue fate classification in stroke. In order to compare objectively the performances of existing deconvolution algorithms and to validate new ones, it is necessary to have access to information on the ground truth after deconvolution. To this end, we developed a numerical simulator of DSC MRI with automatically generated ground truth. This simulator is used to demonstrate the feasability of a full automation of regularization parameters tuning and to establish the robustness of a recent deconvolution algorithm with spatio-temporal regularization. We then propose a new globally convergent deconvolution algorithm. Then, this first part ends with a discussion on another processing step in the DSC-MRI pipeline, the normalisation of the hemodynamic parameters maps extracted from the deconvolved images. In a second part, we work on the prediction of the evolution of the tissue state from longitudinal MRI data. We first demonstrate the interest of modeling longitudinal MRI studies in stroke as a communication channel where information theory provides useful tools to identify the hemodynamic parameters maps carrying the highest predictive information, determine the spatial observation scales providing the optimal predictivity for tissue classification as well as estimate the impact of noise in prediction studies. We then demonstrate the interest of injecting shape descriptors of the ischemic lesion in acute stage in a linear regression model for the prediction of the final infarct volume. We finally propose a classifier of tissue fate based on local binary pattern for the encoding of the spatio-temporal evolution of the perfusion MRI signals