Orientador: Maria de Lourdes Boeno Trindade Galo / Banca: Nilton Nobuhiro Imai / Banca: Francisco Antônio Dupas / Resumo: Nos últimos anos, infestações de plantas aquáticas em reservatórios estão sendo estudadas como um efeito do desequilíbrio causado pela poluição e represamento dos rios. A quantidade excessiva de plantas, conseqüente desse desequilíbrio, dificulta tanto a navegação como a produção de energia elétrica. Esse tipo de ocorrência, assim como a presença de algumas substâncias na água, provocam mudanças na radiância da mesma, registradas por sensores orbitais. Nesse sentido, técnicas de processamento e análise de dados de sensoriamento remoto podem se constituir em uma fonte complementar de dados e fornecer informações relacionadas ao grau de infestação de reservatórios. Nesse contexto, o presente trabalho teve como objetivo verificar a influência da resolução espacial de imagens multiespectrais na detecção e mapeamento de áreas infestadas por plantas aquáticas emersas em um reservatório de pequeno porte, através de utilização de procedimentos de análise multiescala e classificação supervisionada usando redes neurais artificiais. Para isso foram utilizadas imagens IKONOS multiespectrais (4 metros de resolução espacial) do reservatório de Salto Grande localizado na cidade de Americana- SP. Assim, foram geradas imagens multiescala, resultando em imagens de 8, 16 e 32 metros de resolução espacial. Na classificação das imagens, utilizando Redes Neurais Artificiais, os dados de entrada constituíram-se de imagens multiespectrais IKONOS (bandas 1, 2, 3 e 4), imagem de textura (banda do IVP), e uma imagem de índice de vegetação (NDVI). O procedimento metodológico adotado mostrou-se adequado para o mapeamento das variações espectrais da água e detecção das infestações por plantas aquáticas, nos vários níveis de resolução da imagem. Os resultados obtidos mostraram que a classificação pela rede neural, com os parâmetros...(Resumo completo, clicar acesso eletrônico abaixo) / Abstract: In past few years, great infestations of aquatic plants in reservoirs have been studied as an effect of the environmental unbalance caused by pollution and damming of rivers. The excessive amount of plants, deriving from this unbalance, makes navigation and the production of electricity difficult. This kind of occurrence, as well as the appearance of some substances in the water, cause changes in the water radiance detected by satellite sensors. Thus, processing techniques and data analysis may be used as a complementary data source to give information related to the degree of infestation of these plants in reservoirs. So, the present dissertation aimed at verifying the influence of the spatial resolution of multispectral images in the detection and mapping of areas infested by aquatic plants in a small reservoir , through the use of multiscale analysis procedures and supervised classification using artificial neural networks. Multiespectral imagens IKONOS (spatial resolution of 4 meters) of the reservoir of Salto Grande, in the city of Americana-SP were used. So, multiscale images were generated, resulting in images of 8, 16 and 32 meters of spatial resolution. In the classification of these images, using Artificial Neural Networks, the input data was constituted of multispectral images IKONOS (bands 1, 2, 3 and 4), image of texture (band of NIR), and one image of vegetation index (NDVI). The method used was adequate to map the spectral variation of the water and to detect infested areas of aquatic plants in the various levels of resolution of the image. The results obtained showed that the classification by the parameters defined for the original image and applied in the training of the scheme adopted for the different resolution levels was satisfactory. Furthermore, an analysis was made comparing multiscale images classified through crossed comparison, which permits comparing...(Complete abstract click electronic access below) / Mestre

O crescimento descontrolado ocorrido nas atuais metrópoles de países em desenvolvimento requer intensos mapeamentos para a atualização da base de dados geográfica. O intenso processo de urbanização vivido na cidade de São Paulo desde os anos 70 ilustra bem esse cenário. Apesar de existirem levantamentos aéreos e, mais recentemente, imagens de satélite com alta resolução espacial, a necessidade de informações geográficas precisas, rápidas e menos onerosas é, mais do que nunca, um fato. Nesse sentido, a classificação automatizada de imagens de alta resolução espacial tem demonstrado resultados insatisfatórios ao utilizar classificadores pixel a pixel, em especial para áreas urbanas. O crescente sucesso da classificação de imagens baseada em objetos tem estimulado pesquisadores a criar novos meios de superar a limitação das tradicionais técnicas de classificação de imagens. A idéia central da classificação de imagens orientada a objetos é extrair objetos primitivos a partir das imagens e utilizar suas informações para a composição de regras e estratégias a serem aplicadas no processo classificatório. Além da análise espectral, a classificação de imagens baseada em objetos permite envolver análises geométricas e contextuais. Este trabalho reporta o uso da classificação baseada em objetos para detecção da malha viária, aplicado na periferia urbana da cidade de São Paulo. Áreas de ocupação irregular compõem a maior parte da área selecionada para o estudo, sendo que a malha viária reflete bem o padrão de ocupação não planejada dessa região. As ruas são em geral geometricamente irregulares e com diferentes tipos de pavimentação. Detectar a malha viária com base nessas características foi o desafio maior deste trabalho, que teve, como hipótese, a viabilidade do emprego da classificação orientada a objetos para essa finalidade. A metodologia apresentada utiliza uma imagem multiespectral do satélite IKONOS II. Como primeiros passos, processou-se a segmentação e calcularam-se as componentes principais. Classes auxiliares como áreas impermeabilizadas e áreas de solo exposto foram computadas utilizando funções apropriadas. Em suma, a partir das informações geométricas dos objetos, como largura, comprimento, coeficiente de assimetria, área, entre outros, alguns objetos foram selecionados como representantes da malha viária, e então analisados perante a informação contextual, para que fossem classificados como vias pavimentadas e vias não pavimentadas. Os resultados foram analisados mediante três diferentes métodos: 1) inspeção visual, na qual foi analisada qualitativamente a aderência entre as vias extraídas e as vias reais; 2) acurácia da classificação, através de comparações entre a malha viária detectada e a de referência, que forneceu parâmetros estatísticos de qualidade da classificação, como os erros de comissão e omissão ; 3) análise linear comparativa, a qual forneceu parâmetros como integridade (ou completeza) e precisão da malha viária detectada utilizando linhas referenciais e linhas extraídas dos polígonos das vias detectadas, obtidos por morfologia matemática. Considerando o alto grau de heterogeneidade das feições presentes na área de estudo, a acurácia geral alcançada foi boa. Embora a metodologia não tenha produzido um mapa viário, no sentido próprio da palavra, o uso combinado de imagens multispectrais de alta resolução espacial e da classificação baseada em objetos mostrou que a metodologia pode ser utilizada para minerar dados relativos a malha viária e produzir informações significantes para auxiliar a tomada de decisões. / Uncontrolled sprawl occurring in large cities of developing countries requires intensive mapping efforts to update geodatabases. The intense urbanization process experienced since the 70\'s in Sao Paulo city illustrates very well the reported scenario. Despite aerial data and, more recent, high spatial resolution satellite data which have been employed as basis for mapping, the need for precise, faster and cheaper mapping efforts is real. In this sense, automated classification of high resolution imagery has demonstrated unsatisfactory results when traditional per-pixel classifiers are used, especially for urban areas. The increasing success of object-based classification has stimulated researchers to create new methodologies to overcome this shortcoming of traditional approaches. The object-based image classification\'s idea is extract object-primitives from images and then use their information to compose rules and strategies to be applied on the classification process. Beyond the spectral analysis, geometric, and contextual analysis are also addressed on object-based classification. This work reports the use of object-based image classification applied on road detection over the suburban area of Sao Paulo city. Informal settlements compose the most part of the study area and the transportation network reflects the unplanned occupation. Roads are geometrically irregular and with different kind of pavements. Detecting roads based on these characteristics was the biggest challenge faced here, and this work hypothesizes object-based classification can be used to. The methodology presented employs an IKONOS II data. At first, principal components and segmentation were computed and then auxiliary data for impervious surface and bare soil areas were previously calculated from customized features. In short, based on geometric information as width, length, asymmetry, area, and more, objects were elected as road and then analyzed through contextual information as paved road or unpaved road. Results were analyzed under three different ways: 1) visual inspection, where the adherence between extracted road and real ones provided a good indicator for qualitative analysis ; 2) classification accuracy, by comparing detected road areas and referential ones, which provided statistical parameters for quality as omission and commission error ; 3) linear comparative analysis, which provided parameters as correctness and completeness using referential lines and lines arose from extracted areas based on mathematical morphology tools. Regarding the high degree of heterogeneity of features present on study area, the overall accuracy reached is good. Despite the methodology did not produce a road map, the results shown the combined use of high resolution multi-spectral imagery and object-based classification can effectively mine road features, producing significant information to support decision makers.

Cada imagem pode ser representada como uma combinação de diversas características, como por exemplo o histograma de intensidades de cor ou propriedades de textura da imagem. Essas características compõem um vetor multidimensional que representa a imagem. É comum esse vetor ser dado como entrada para um método de classificação de padrões que, após aprender por meio de diversos exemplos, pode gerar um modelo de decisão. Estudos sugerem evidências de que a preparação das imagens-- por meio da especificação cuidadosa da aquisição, pré-processamento e segmentação-- pode impactar significativamente a classificação. Além da falta de tratamento das imagens antes da extração de características, o desbalanceamento de classes também se apresenta como um obstáculo para que a classificação seja satisfatória. Imagens possuem características que podem ser exploradas para melhorar a descrição dos objetos de interesse e, portanto, sua classificação. Entre as possibilidades de melhorias estão: a redução do número de intensidades das imagens antes da extração de características ao invés de métodos de quantização no vetor já extraído; e a geração de imagens a partir das originais, de forma a promover o balanceamento de bases de dados cujo número de exemplos de cada classe é desbalanceado. Portanto, a proposta desta dissertação é melhorar a classificação de imagens utilizando métodos de processamento de imagens antes da extração de características. Especificamente, busca analisar a influência do balanceamento de bases de dados e da quantização na classificação. Este estudo analisa ainda a visualização do espaço de características após os métodos de geração artificial de imagens e de interpolação das características extraídas das imagens originais (SMOTE), comparando como espaço original. A ênfase dessa visualização se dá na observação da importância do rebalanceamento das classes. Os resultados obtidos indicam que a quantização simplifica as imagens antes da extração de características e posterior redução de dimensionalidade, produzindo vetores mais compactos; e que o rebalanceamento de classes de imagens através da geração de imagens artificiais pode melhorar a classificação da base de imagens, em relação à classificação original e ao uso de métodos no espaço de características já extraídas. / Each image can be represented by a combination of several features like color frequency and texture properties. Those features compose a multidimensional vector, which represents the original image. Commonly this vector is given as an input to a classification method that can learn from examplesand build a decision model. The literature suggests that image preparation steps like acute acquisition, preprocessing and segmentation can positively impact such classification. Besides that, class unbalancing is also a barrier to achieve good classification accuracy. Some features and methods can be explored to improveobjects\' description, thus their classification. Possible suggestions include: reducing colors number before feature extraction instead of applying quantization methods to raw vectors already extracted; and generating synthetic images from original ones, to balance the number of samples in an uneven data set. We propose to improve image classification using image processing methods before feature extraction. Specifically we want to analyze the influence of both balancing and quantization methods while applied to datasets in a classification routine. This research also analyses the visualization of feature space after the artificial image generation and feature interpolation (SMOTE), against to original space. Such visualization is used because it allows us to know how important is the rebalacing method. The results show that quantization simplifies imagesby producing compacted vectors before feature extraction and dimensionality reduction; and that using artificial generation to rebalance image datasets can improve classification, when compared to the original one and to applying methods on the already extracted feature vectors.

A modelagem hidrológica quando associada aos recursos do sensoriamento remoto e do geoprocessamento torna-se uma ferramenta importante, pois é capaz de estabelecer diferentes cenários da cobertura e do uso da terra e suas implicações na drenagem urbana, auxiliando no planejamento urbano. Entretanto, a relação entre o modelo chuva x vazão e tais técnicas, com finalidade de avaliar classificadores de imagens a partir de hidrogramas de cheia não foi encontrada na literatura, tornando esse o objetivo principal desta tese. Para isso, foram utilizadas três imagens de satélite de diferentes resoluções espaciais (0,5m, 5m e 15m) e três algoritmos classificadores (Máxima Verossimilhança, Máquinas Vetores Suporte e Análise Orientada a Objeto) e formados conjuntos denominado \"classificador-imagem\" para classificação da cobertura e do uso da terra. As áreas das classes dos usos da terra de cada conjunto \"classificador-imagem\" e os valores de Curve Number foram os principais dados de entrada do modelo chuva-vazão NRCS, que permitiu gerar os hidrogramas de cheia para cada caso. Os hidrogramas simulados foram comparados aos hidrogramas observados na bacia e avaliados, quanto a sua representatividade, pelo coeficiente de Nash Sutcliffe. As classificações do uso da terra foram avaliadas pelo Índice Kappa, com valores de 0,58 a 0,99 e pela Exatidão Global, com valores de 0,64 a 0,99. Para as vazões, o coeficiente de Nash Sutcliffe foi considerado satisfatório (NS<0,50) em duas simulações e, nas demais simulações, considerado muito bom (NS>0,75). Para fornecer subsídio a tomada de decisão, foi realizada uma análise multicritério dos conjuntos classificador-imagem, que permitiu classificar os conjuntos com maior desempenho: 1°) o classificador SVM e a imagem Landsat-8; 2°) o classificador MaxVer e a imagem WordView-II; 3°) o classificador NN e a imagem RapidEye. / Hydrological modeling when associated with remote sensing and geoprocessing resources becomes an important tool, because it is able to establish different land use scenarios and its implications for urban drainage, assisting in urban planning. However, the relationship between the routing model and such techniques, for purpose to evaluate images classifiers from the runoff hydrograph was not found in the literature, making this the main objective of this thesis. Thereunto, three satellite images were used in different spatial resolutions (0.5m, 5m and 15m) and three algorithms classifiers (Maximum Likelihood, Support Vector Machine and Oriented Object Analysis) and composed sets called \"classifier-image\" for the land use classification. The areas of the land use classes of each set \"classifier-image\" and the Curve Number values were the main input of the routing model NRCS, which allowed generating the runoff hydrograph for each case. The simulated hydrographs were compared to the observed hydrograph in the basin and evaluated their representativeness through the Nash Sutcliffe coefficient. Kappa Index was calculated to evaluate land use classifications, with values between 0.58 to 0.99 and Global accuracy between 0.64 to 0.99. Towards the flows rates, the Nash Sutcliffe coefficient was considered satisfactory for two simulations (NS<0,50) and, to other simulations, considered very good (NS>0,75). To provide subsidy to decision-making, it carried out a multi-criteria analysis of the classifier-image sets, that allowed to classify the set with higher performance: 1) SVM classifier and Landsat-8 image; 2) MaxVer classifier and WorldView-II image; 3) NN classifier and RapidEye image.

Cette thèse a pour objectif de proposer un système de reconnaissance d’images utilisant des informations attentionnelles. Nous nous intéressons à la capacité d’une telle approche à améliorer la complexité en temps de calcul et en utilisation mémoire pour la reconnaissance d’objets. Dans un premier temps, nous avons proposé d’utiliser un système d’attention visuelle comme filtre pour réduire le nombre de points d’intérêt générés par les détecteurs traditionnels [Awad 12]. En utilisant l’architecture attentionnelle proposée par Perreira da Silva comme filtre [Awad 12] sur la base d’images de VOC 2005, nous avons montré qu’un filtrage de 60% des points d’intérêt (extraits par Harris-Laplace et Laplacien) ne fait diminuer que légèrement la performance d’un système de reconnaissance d’objets (différence moyenne de AUC ~ 1%) alors que le gain en complexité est important (40% de gain en vitesse de calcul et 60% en complexité). Par la suite, nous avons proposé un descripteur hybride perceptuel-texture [Awad 14] qui caractérise les informations fréquentielles de certaines caractéristiques considérées comme perceptuellement intéressantes dans le domaine de l’attention visuelle, comme la couleur, le contraste ou l’orientation. Notre descripteur a l’avantage de fournir des vecteurs de caractéristiques ayant une dimension deux fois moindre que celle des descripteurs proposés dans l’état de l’art. L’expérimentation de ce descripteur sur un système de reconnaissance d’objets (le détecteur restant SIFT), sur la base d’images de VOC 2007, a montré une légère baisse de performance (différence moyenne de précision ~5%) par rapport à l’algorithme original, basé sur SIFT mais gain de 50% en complexité. Pour aller encore plus loin, nous avons proposé une autre expérimentation permettant de tester l’efficacité globale de notre descripteur en utilisant cette fois le système d’attention visuelle comme détecteur des points d’intérêt sur la base d’images de VOC 2005. Là encore, le système n’a montré qu’une légère baisse de performance (différence moyenne de précision ~3%) alors que la complexité est réduite de manière drastique (environ 50% de gain en temps de calcul et 70% en complexité). / The main objective of this thesis is to propose a pipeline for an object recognition algorithm, near to human perception, and at the same time, address the problems of Content Based image retrieval (CBIR) algorithm complexity : query run time and memory allocation. In this context, we propose a filter based on visual attention system to select salient points according to human interests from the interest points extracted by a traditionnal interest points detectors. The test of our approach, using Perreira Da Silva’s system as filter, on VOC 2005 databases, demonstrated that we can maintain approximately the same performance of a object recognition system by selecting only 40% of interest points (extracted by Harris-Laplace and Laplacian), while having an important gain in complexity (40% gain in query-run time and 60% in complexity). Furthermore, we address the problem of high dimensionality of descriptor in object recognition system. We proposed a new hybrid texture descriptor, representing the spatial frequency of some perceptual features extracted by a visual attention system. This descriptor has the advantage of being lower dimension vs. traditional descriptors. Evaluating our descriptor with an object recognition system (interest points detectors are Harris-Laplace & Laplacian) on VOC 2007 databases showed a slightly decrease in the performance (with 5% loss in Average Precision) compared to the original system, based on SIFT descriptor (with 50% complexity gain). In addition, we evaluated our descriptor using a visual attention system as interest point detector, on VOC 2005 databases. The experiment showed a slightly decrease in performance (with 3% loss in performance), meanwhile we reduced drastically the complexity of the system (with 50% gain in run-query time and 70% in complexity).

L'augmentation exponentielle du nombre d'images nécessite des moyens efficaces pour les classer en fonction de leur contenu visuel. Le sac de mot visuel (Bag-Of-visual-Words, BOW), en raison de sa simplicité et de sa robustesse, devient l'approche la plus populaire. Malheureusement, cette approche ne prend pas en compte de l'information spatiale, ce qui joue un rôle important dans les catégories de modélisation d'image. Récemment, Lazebnik ont introduit la représentation pyramidale spatiale (Spatial Pyramid Representation, SPR) qui a incorporé avec succès l'information spatiale dans le modèle BOW. Néanmoins, ce système de correspondance rigide empêche la SPR de gérer les variations et les transformations d'image. L'objectif principal de cette thèse est d'étudier un modèle de chaîne de correspondance plus souple qui prend l'avantage d'histogrammes de BOW locaux et se rapproche de la correspondance de la chaîne. Notre première contribution est basée sur une représentation en chaîne et une nouvelle distance d'édition (String Matching Distance, SMD) bien adapté pour les chaînes de l'histogramme qui peut calculer efficacement par programmation dynamique. Un noyau d'édition correspondant comprenant à la fois d'une pondération et d'un système pyramidal est également dérivée. La seconde contribution est une version étendue de SMD qui remplace les opérations d'insertion et de suppression par les opérations de fusion entre les symboles successifs, ce qui apporte de la souplesse labours et correspond aux images. Toutes les distances proposées sont évaluées sur plusieurs jeux de données tâche de classification et sont comparés avec plusieurs approches concurrentes / The exponential increasing of the number of images requires efficient ways to classify them based on their visual content. The most successful and popular approach is the Bag of visual Word (BoW) representation due to its simplicity and robustness. Unfortunately, this approach fails to capture the spatial image layout, which plays an important roles in modeling image categories. Recently, Lazebnik et al (2006) introduced the Spatial Pyramid Representation (SPR) which successfully incorporated spatial information into the BoW model. The idea of their approach is to split the image into a pyramidal grid and to represent each grid cell as a BoW. Assuming that images belonging to the same class have similar spatial distributions, it is possible to use a pairwise matching as similarity measurement. However, this rigid matching scheme prevents SPR to cope with image variations and transformations. The main objective of this dissertation is to study a more flexible string matching model. Keeping the idea of local BoW histograms, we introduce a new class of edit distance to compare strings of local histograms. Our first contribution is a string based image representation model and a new edit distance (called SMD for String Matching Distance) well suited for strings composed of symbols which are local BoWs. The new distance benefits from an efficient Dynamic Programming algorithm. A corresponding edit kernel including both a weighting and a pyramidal scheme is also derived. The performance is evaluated on classification tasks and compared to the standard method and several related methods. The new method outperforms other methods thanks to its ability to detect and ignore identical successive regions inside images. Our second contribution is to propose an extended version of SMD replacing insertion and deletion operations by merging operations between successive symbols. In this approach, the number of sub regions ie. the grid divisions may vary according to the visual content. We describe two algorithms to compute this merge-based distance. The first one is a greedy version which is efficient but can produce a non optimal edit script. The other one is an optimal version but it requires a 4th degree polynomial complexity. All the proposed distances are evaluated on several datasets and are shown to outperform comparable existing methods.

Indianapolis / E-scooters are ubiquitous and their number keeps escalating, increasing their interactions with other vehicles on the road. E-scooter riders have an atypical behavior that varies enormously from other vulnerable road users, creating new challenges for vehicle active safety systems and automated driving functionalities. The detection of e-scooter riders by other vehicles is the first step in taking care of the risks. This research presents a novel vision-based system to differentiate between e-scooter riders and regular pedestrians and a benchmark dataset for e-scooter riders in natural environments. An efficient system pipeline built using two existing state-of-the-art convolutional neural networks (CNN), You Only Look Once (YOLOv3) and MobileNetV2, performs detection of these vulnerable e-scooter riders.

Labeling of trademark images with Vienna codes from the Vienna classification is a manual process carried out by domain experts, which enables searching trademark image databases using specific keywords that describe the semantic meaning of the figurative elements. In this research, we are investigating how application of supervised learning algorithms can improve and automate the manual process of labeling of new un-labeled trademark images. The successful implementation of deep learning algorithms in the task of computer vision for image classification has motivated us to investigate which of the supervised learning algorithms performs better trademark image classification. More specifically, to solve the problem of identification of figurative elements in new un-labeled images, we have used multi-class image classification approach based on deep learning and machine learning. To address this problem, we have generated a unique benchmarking dataset composed of 14,500 unique logos extracted from the European Union Intellectual Property Office Open Data Portal. The results after executing a set of controlled experiments on the given dataset indicate that deep learning models have overall better performance than machine learning models. In particular, CNN models reach better accuracy and precision, and significantly higher recall and F1 score for shorter training times, compared to recurrent neural networks such as LSTMs and GRUs. From the machine learning models, results indicate that Support Vector Machines have higher accuracy and overall better performance time compared to Decision Trees, Random Forests and Naïve Bayes models. This study shows that deep learning models can solve the problem of the labeling of trademark images with Vienna codes, and that can be applied by Intellectual Property Offices in real-world application for automation of the classification task which is carried out manually by the domain experts.

The subject of this thesis is a method proposed for automated evaluation of the parameters of X-ray of cystic disorders in human jawbones. The main problem in medical diagnostic is the low repeatability due to the subjective evaluation of images without using a tool for image processing. In this thesis are described the basic steps of image processing, various methods of image segmentation and chosen segmentation method live-wire. Selected segments were processed in the ImageJ Java environment. In the cystic regions their basic statistical and shape properties were evaluated. The obtained values were used for learning the classification model (decision tree) in the environment RapidMiner. This model was used to create a plug-in for automatic classification of the type of cysts in the program ImageJ.

Cette thèse s'inscrit dans le contexte de diagnostic assisté par ordinateur pour la localisation subcellulaire des protéines dans les images microscopiques. L'objectif est la conception et le développement d'un système de classification automatique permettant d'identifier le compartiment cellulaire dans lequel une protéine d'intérêt exerce son activité biologique. Afin de surmonter les difficultés rencontrées pour discerner les compartiments cellulaires présents dans les images microscopiques, les systèmes décrits dans la littérature proposent d'extraire plusieurs descripteurs associés à une combinaison de classifieurs. Dans cette thèse, nous proposons un schéma de classification différent répondant mieux aux besoins de généricité et de flexibilité pour traiter différentes bases d'images.Dans le but de fournir une caractérisation riche des images microscopiques, nous proposons un nouveau système de représentation permettant d'englober de multiples descripteurs visuels identifiés dans les différentes approches d'extraction de caractéristiques : locale, fréquentielle, globale et par région. Nous formulons ensuite le problème de fusion et de sélection des caractéristiques sous forme d'un problème de sélection de noyaux. Basé sur l'apprentissage de noyaux multiples (MKL), les tâches de sélection et de fusion de caractéristiques sont considérées simultanément. Les expériences effectuées montrent que la plateforme de classification proposée est à la fois plus simple, plus générique et souvent plus performante que les autres approches de la littérature. Dans le but d'approfondir notre étude sur l'apprentissage de noyaux multiples, nous définissons un nouveau formalisme d'apprentissage MKL réalisé en deux étapes. Cette contribution consiste à proposer trois termes régularisant liés à la résolution du problème d'apprentissage des poids associés à une combinaison linéaire de noyaux, problème reformulé en un problème de classification à vaste marge dans l'espace des couples. Le premier terme régularisant proposé assure une sélection parcimonieuse des noyaux. Les deux autres termes ont été conçus afin de tenir compte de la similarité entre les noyaux via une métrique basée sur la corrélation. Les différentes expérimentations réalisées montrent que le formalisme proposé permet d'obtenir des résultats de même ordre que les méthodes de référence, mais offrant l'avantage d'utiliser moins de fonctions noyaux. / This thesis arises in the context of computer aided analysis for subcellular protein localization in microscopic images. The aim is the establishment of an automatic classification system allowing to identify the cellular compartment in which a protein of interest exerts its biological activity. In order to overcome the difficulties in attempting to discern the cellular compartments in microscopic images, the existing state-of-art systems use several descriptors to train an ensemble of classifiers. In this thesis, we propose a different classification scheme wich better cope with the requirement of genericity and flexibility to treat various image datasets. Aiming to provide an efficient image characterization of microscopic images, a new feature system combining local, frequency-domain, global, and region-based features is proposed. Then, we formulate the problem of heterogeneous feature fusion as a kernel selection problem. Using multiple kernel learning, the problems of optimal feature sets selection and classifier training are simultaneously resolved. The proposed combination scheme leads to a simple and a generic framework capable of providing a high performance for microscopy image classification. Extensive experiments were carried out using widely-used and best known datasets. When compared with the state-of-the-art systems, our framework is more generic and outperforms other classification systems. To further expand our study on multiple kernel learning, we introduce a new formalism for learning with multiple kernels performed in two steps. This contribution consists in proposing three regularized terms with in the minimization of kernels weights problem, formulated as a classification problem using Separators with Vast Margin on the space of pairs of data. The first term ensures that kernels selection leads to a sparse representation. While the second and the third terms introduce the concept of kernels similarity by using a correlation measure. Experiments on various biomedical image datasets show a promising performance of our method compared to states of art methods.