Il y a un intérêt grandissant pour le développement d’outils de traitements adaptés aux images multimodales (plusieurs images de la même scène acquises avec différentes caractéristiques). Permettant une représentation plus complète de la scène, ces images multimodales ont de l'intérêt dans plusieurs domaines du traitement d'images, mais les exploiter et les manipuler de manière optimale soulève plusieurs questions. Cette thèse étend les représentations hiérarchiques, outil puissant pour le traitement et l’analyse d’images classiques, aux images multimodales afin de mieux exploiter l’information additionnelle apportée par la multimodalité et améliorer les techniques classiques de traitement d’images. Cette thèse se concentre sur trois différentes multimodalités fréquemment rencontrées dans le domaine de la télédétection. Nous examinons premièrement l’information spectrale-spatiale des images hyperspectrales. Une construction et un traitement adaptés de la représentation hiérarchique nous permettent de produire une carte de segmentation de l'image optimale vis-à-vis de l'opération de démélange spectrale. Nous nous concentrons ensuite sur la multimodalité temporelle, traitant des séquences d’images hyperspectrales. En utilisant les représentations hiérarchiques des différentes images de la séquence, nous proposons une nouvelle méthode pour effectuer du suivi d’objet et l’appliquons au suivi de nuages de gaz chimique dans des séquences d’images hyperspectrales dans le domaine thermique infrarouge. Finalement, nous étudions la multimodalité sensorielle, c’est-à-dire les images acquises par différents capteurs. Nous appuyant sur le concept des tresses de partitions, nous proposons une nouvelle méthodologie de segmentation se basant sur un cadre de minimisation d’énergie. / There is a growing interest in the development of adapted processing tools for multimodal images (several images acquired over the same scene with different characteristics). Allowing a more complete description of the scene, multimodal images are of interest in various image processing fields, but their optimal handling and exploitation raise several issues. This thesis extends hierarchical representations, a powerful tool for classical image analysis and processing, to multimodal images in order to better exploit the additional information brought by the multimodality and improve classical image processing techniques. %when applied to real applications. This thesis focuses on three different multimodalities frequently encountered in the remote sensing field. We first investigate the spectral-spatial information of hyperspectral images. Based on an adapted construction and processing of the hierarchical representation, we derive a segmentation which is optimal with respect to the spectral unmixing operation. We then focus on the temporal multimodality and sequences of hyperspectral images. Using the hierarchical representation of the frames in the sequence, we propose a new method to achieve object tracking and apply it to chemical gas plume tracking in thermal infrared hyperspectral video sequences. Finally, we study the sensorial multimodality, being images acquired with different sensors. Relying on the concept of braids of partitions, we propose a novel methodology of image segmentation, based on an energetic minimization framework.

Nous étudions la problématique de classification dite fine en se concentrant sur la détermination des espèces botaniques à partir d’images de feuilles. Nous nous intéressons aussi bien à la description et la représentation de l’objet qu’aux algorithmes de classification et des scénarios d’identification utiles à l’utilisateur. Nous nous inspirons du processus manuel des botanistes pour introduire une nouvelle représentation hiérarchique des feuilles. Nous proposons aussi un nouveau mécanisme permettant d’attirer l’attention au tour de certains points caractéristiques de l’objet et d’apprendre des signatures spécifiques à chaque catégorie.Nous adoptons une stratégie de classification hiérarchique utilisant une série de classifieurs locaux allant des plus grossiers vers les plus fins; la classification locale étant basée sur des rapports de vraisemblance. L’algorithme fournit une liste d’estimations ordonnées selon leurs rapports de vraisemblance. Motivés par les applications, nous introduisons un autre scénario proposant à l’utilisateur un ensemble de confiance contenant la bonne espèce avec une probabilité très élevée. Un nouveau critère de performance est donc considéré : la taille de l’ensemble retourné. Nous proposons un modèle probabiliste permettant de produire de tels ensembles de confiance. Toutes les méthodes sont illustrées sur plusieurs bases de feuilles ainsi que des comparaisons avec les méthodes existantes. / We introduce models for fine-grained categorization, focusing on determining botanical species from leaf images. Images with both uniform and cluttered background are considered and several identification scenarios are presented, including different levels of human participation. Both feature extraction and classification algorithms are investigated. We first leverage domain knowledge from botany to build a hierarchical representation of leaves based on IdKeys, which encode invariable characteristics, and refer to geometric properties (i.e., landmarks) and groups of species (e.g., taxonomic categories). The main idea is to sequentially refine the object description and thus narrow down the set of candidates during the identification task. We also introduce vantage feature frames as a more generic object representation and a mechanism for focusing attention around several vantage points (where to look) and learning dedicated features (what to compute). Based on an underlying coarse-to-fine hierarchy, categorization then proceeds from coarse-grained to fine-grained using local classifiers which are based on likelihood ratios. Motivated by applications, we also introduce on a new approach and performance criterion: report a subset of species whose expected size is minimized subject to containing the true species with high probability. The approach is model-based and outputs a confidence set in analogy with confidence intervals in classical statistics. All methods are illustrated on multiple leaf datasets with comparisons to existing methods.

In this dissertation, we introduce a novel text representation method mainly used for text classification purpose. The presented representation method is initially based on a variety of closeness relationships between pairs of words in text passages within the entire corpus. This representation is then used as the basis for our multi-level lightweight ontological representation method (TOR-FUSE), in which documents are represented based on their contexts and the goal of the learning task. The method is unlike the traditional representation methods, in which all the documents are represented solely based on the constituent words of the documents, and are totally isolated from the goal that they are represented for. We believe choosing the correct granularity of representation features is an important aspect of text classification. Interpreting data in a more general dimensional space, with fewer dimensions, can convey more discriminative knowledge and decrease the level of learning perplexity. The multi-level model allows data interpretation in a more conceptual space, rather than only containing scattered words occurring in texts. It aims to perform the extraction of the knowledge tailored for the classification task by automatic creation of a lightweight ontological hierarchy of representations. In the last step, we will train a tailored ensemble learner over a stack of representations at different conceptual granularities. The final result is a mapping and a weighting of the targeted concept of the original learning task, over a stack of representations and granular conceptual elements of its different levels (hierarchical mapping instead of linear mapping over a vector). Finally the entire algorithm is applied to a variety of general text classification tasks, and the performance is evaluated in comparison with well-known algorithms.

In this dissertation, we introduce a novel text representation method mainly used for text classification purpose. The presented representation method is initially based on a variety of closeness relationships between pairs of words in text passages within the entire corpus. This representation is then used as the basis for our multi-level lightweight ontological representation method (TOR-FUSE), in which documents are represented based on their contexts and the goal of the learning task. The method is unlike the traditional representation methods, in which all the documents are represented solely based on the constituent words of the documents, and are totally isolated from the goal that they are represented for. We believe choosing the correct granularity of representation features is an important aspect of text classification. Interpreting data in a more general dimensional space, with fewer dimensions, can convey more discriminative knowledge and decrease the level of learning perplexity. The multi-level model allows data interpretation in a more conceptual space, rather than only containing scattered words occurring in texts. It aims to perform the extraction of the knowledge tailored for the classification task by automatic creation of a lightweight ontological hierarchy of representations. In the last step, we will train a tailored ensemble learner over a stack of representations at different conceptual granularities. The final result is a mapping and a weighting of the targeted concept of the original learning task, over a stack of representations and granular conceptual elements of its different levels (hierarchical mapping instead of linear mapping over a vector). Finally the entire algorithm is applied to a variety of general text classification tasks, and the performance is evaluated in comparison with well-known algorithms.

La finesse de la résolution spectrale des images hyperspectrales en télédétection permet une analyse précise de la scène observée, mais leur résolution spatiale est limitée, et un pixel acquis par le capteur est souvent un mélange des contributions de différents matériaux. Le démélange spectral permet d'estimer les spectres des matériaux purs (endmembers) de la scène, et leurs abondances dans chaque pixel. Les endmembers sont souvent supposés être parfaitement représentés par un seul spectre, une hypothèse fausse en pratique, chaque matériau ayant une variabilité intra-classe non négligeable. Le but de cette thèse est de développer des algorithmes prenant mieux en compte ce phénomène. Nous effectuons le démélange localement, dans des régions bien choisies de l'image où les effets de la variabilité sont moindres, en éliminant automatiquement les endmembers non pertinents grâce à de la parcimonie collaborative. Dans une autre approche, nous raffinons l'estimation des abondances en utilisant la structure de groupe d'un dictionnaire d'endmembers extrait depuis les données. Ensuite, nous proposons un modèle de mélange linéaire étendu, basé sur des considérations physiques, qui modélise la variabilité spectrale par des facteurs d'échelle, et développons des algorithmes d'optimisation pour en estimer les paramètres. Ce modèle donne des résultats facilement interprétables et de meilleures performances que d'autres approches de la littérature. Nous étudions enfin deux applications de ce modèle pour confirmer sa pertinence. / The fine spectral resolution of hyperspectral remote sensing images allows an accurate analysis of the imaged scene, but due to their limited spatial resolution, a pixel acquired by the sensor is often a mixture of the contributions of several materials. Spectral unmixing aims at estimating the spectra of the pure materials (called endmembers) in the scene, and their abundances in each pixel. The endmembers are usually assumed to be perfectly represented by a single spectrum, which is wrong in practice since each material exhibits a significant intra-class variability. This thesis aims at designing unmixing algorithms to better handle this phenomenon. First, we perform the unmixing locally in well chosen regions of the image where variability effects are less important, and automatically discard wrongly estimated local endmembers using collaborative sparsity. In another approach, we refine the abundance estimation of the materials by taking into account the group structure of an image-derived endmember dictionary. Second, we introduce an extended linear mixing model, based on physical considerations, modeling spectral variability in the form of scaling factors, and develop optimization algorithms to estimate its parameters. This model provides easily interpretable results and outperforms other state-of-the-art approaches. We finally investigate two applications of this model to confirm its relevance.

In this dissertation, we introduce a novel text representation method mainly used for text classification purpose. The presented representation method is initially based on a variety of closeness relationships between pairs of words in text passages within the entire corpus. This representation is then used as the basis for our multi-level lightweight ontological representation method (TOR-FUSE), in which documents are represented based on their contexts and the goal of the learning task. The method is unlike the traditional representation methods, in which all the documents are represented solely based on the constituent words of the documents, and are totally isolated from the goal that they are represented for. We believe choosing the correct granularity of representation features is an important aspect of text classification. Interpreting data in a more general dimensional space, with fewer dimensions, can convey more discriminative knowledge and decrease the level of learning perplexity. The multi-level model allows data interpretation in a more conceptual space, rather than only containing scattered words occurring in texts. It aims to perform the extraction of the knowledge tailored for the classification task by automatic creation of a lightweight ontological hierarchy of representations. In the last step, we will train a tailored ensemble learner over a stack of representations at different conceptual granularities. The final result is a mapping and a weighting of the targeted concept of the original learning task, over a stack of representations and granular conceptual elements of its different levels (hierarchical mapping instead of linear mapping over a vector). Finally the entire algorithm is applied to a variety of general text classification tasks, and the performance is evaluated in comparison with well-known algorithms.