Selecionar atributos é, por vezes, uma atividade necessária para o correto desenvolvimento de tarefas de aprendizado de máquina. Em Mineração de Textos, reduzir o número de atributos em uma base de textos é essencial para a eficácia do processo e a compreensibilidade do conhecimento extraído, uma vez que se lida com espaços de alta dimensionalidade e esparsos. Quando se lida com contextos nos quais a coleção de textos é não-rotulada, métodos não-supervisionados de redução de atributos são utilizados. No entanto, não existe forma geral predefinida para a obtenção de medidas de utilidade de atributos em métodos não-supervisionados, demandando um esforço maior em sua realização. Assim, este trabalho aborda a seleção não-supervisionada de atributos por meio de um estudo exploratório de métodos dessa natureza, comparando a eficácia de cada um deles na redução do número de atributos em aplicações de Mineração de Textos. Dez métodos são comparados - Ranking porTerm Frequency, Ranking por Document Frequency, Term Frequency-Inverse Document Frequency, Term Contribution, Term Variance, Term Variance Quality, Método de Luhn, Método LuhnDF, Método de Salton e Zone-Scored Term Frequency - sendo dois deles aqui propostos - Método LuhnDF e Zone-Scored Term Frequency. A avaliação se dá em dois focos, supervisionado, pelo medida de acurácia de quatro classificadores (C4.5, SVM, KNN e Naïve Bayes), e não-supervisionado, por meio da medida estatística de Expected Mutual Information Measure. Aos resultados de avaliação, aplica-se o teste estatístico de Kruskal-Wallis para determinação de significância estatística na diferença de desempenho dos diferentes métodos de seleção de atributos comparados. Seis bases de textos são utilizadas nas avaliações experimentais, cada uma relativa a um grande domínio e contendo subdomínios, os quais correspondiam às classes usadas para avaliação supervisionada. Com esse estudo, este trabalho visa contribuir com uma aplicação de Mineração de Textos que visa extrair taxonomias de tópicos a partir de bases textuais não-rotuladas, selecionando os atributos mais representativos em uma coleção de textos. Os resultados das avaliações mostram que não há diferença estatística significativa entre os métodos não-supervisionados de seleção de atributos comparados. Além disso, comparações desses métodos não-supervisionados com outros supervisionados (Razão de Ganho e Ganho de Informação) apontam que é possível utilizar os métodos não-supervisionados em atividades supervisionadas de Mineração de Textos, obtendo eficiência compatível com os métodos supervisionados, dado que não detectou-se diferença estatística nessas comparações, e com um custo computacional menor / Feature selection is an activity sometimes necessary to obtain good results in machine learning tasks. In Text Mining, reducing the number of features in a text base is essential for the effectiveness of the process and the comprehensibility of the extracted knowledge, since it deals with high dimensionalities and sparse contexts. When dealing with contexts in which the text collection is not labeled, unsupervised methods for feature reduction have to be used. However, there aren\'t any general predefined feature quality measures for unsupervised methods, therefore demanding a higher effort for its execution. So, this work broaches the unsupervised feature selection through an exploratory study of methods of this kind, comparing their efficacies in the reduction of the number of features in the Text Mining process. Ten methods are compared - Ranking by Term Frequency, Ranking by Document Frequency, Term Frequency-Inverse Document Frequency, Term Contribution, Term Variance, Term Variance Quality, Luhn\'s Method, LuhnDF Method, Salton\'s Method and Zone-Scored Term Frequency - and two of them are proposed in this work - LuhnDF Method and Zone-Scored Term Frequency. The evaluation process is done in two ways, supervised, through the accuracy measure of four classifiers (C4.5, SVM, KNN and Naïve Bayes), and unsupervised, using the Expected Mutual Information Measure. The evaluation results are submitted to the statistical test of Kruskal-Wallis in order to determine the statistical significance of the performance difference of the different feature selection methods. Six text bases are used in the experimental evaluation, each one related to one domain and containing sub domains, which correspond to the classes used for supervised evaluation. Through this study, this work aims to contribute with a Text Mining application that extracts topic taxonomies from unlabeled text collections, through the selection of the most representative features in a text collection. The evaluation results show that there is no statistical difference between the unsupervised feature selection methods compared. Moreover, comparisons of these unsupervised methods with other supervised ones (Gain Ratio and Information Gain) show that it is possible to use unsupervised methods in supervised Text Mining activities, obtaining an efficiency compatible with supervised methods, since there isn\'t any statistical difference the statistical test detected in these comparisons, and with a lower computational effort

Há muitos anos, os sistemas de comitê já tem se mostrado um método eficiente para aumentar a acurácia e estabilidade de algoritmos de aprendizado nas décadas recentes, embora sua construção tem uma questão para ser elucidada: diversidade. O desacordo entre os modelos que compõe o comitê pode ser gerado quando eles são contruídos sob diferentes circunstâncias, tais como conjunto de dados de treinamento, configuração dos parâmetros e a seleção dos algoritmos de aprendizado. O ensemble pode ser visto como uma estrutura com três níveis: espaço de entrada, a base de componentes e o bloco de combinação das respostas dos componentes. Neste trabalho é proposta uma abordagem multi-nível usando Algoritmos Genéticos para construir um ensemble de Máquinas de Vetor de Suporte por Mínimos Quadrados ou LS-SVM, realizando uma seleção de atributos no espaço de entrada, parametrização e a escolha de quais modelos irão compor o comitê no nível de componentes e a busca por um vetor de pesos que melhor represente a importância de cada classificador na resposta final do comitê. De forma a avaliar a performance da abordagem proposta, foram utilizados alguns benchmarks do repositório da UCI para comparar com outros algoritmos de classificação. Além disso, também foram comparados os resultados da abordagem proposta com métodos de aprendizagem profunda nas bases de dados MNIST e CIFAR e se mostraram bastante satisfatórios. / Many years ago, the ensemble systems have been shown to be an efficient method to increase the accuracy and stability of learning algorithms in recent decades, although its construction has a question to be elucidated: diversity. The disagreement among the models that compose the ensemble can be generated when they are built under different circumstances, such as training dataset, parameter setting and selection of learning algorithms. The ensemble may be viewed as a structure with three levels: input space, the base components and the combining block of the components responses. In this work is proposed a multi-level approach using genetic algorithms to build the ensemble of Least Squares Support Vector Machines (LS-SVM), performing a feature selection in the input space, the parameterization and the choice of which models will compose the ensemble at the component level and finding a weight vector which best represents the importance of each classifier in the final response of the ensemble. In order to evaluate the performance of the proposed approach, some benchmarks from UCI Repository have been used to compare with other classification algorithms. Also, the results obtained by our approach were compared with some deep learning methods on the datasets MNIST and CIFAR and proved very satisfactory.

In silico experimentation is the process of using computational and statistical models to predict medicinal properties in chemicals; as a means of reducing lab work and increasing success rate this process has become an important part of modern drug development. There are various ways of representing molecules - the problem that motivated this paper derives from collecting substructures of the chemical into what is known as fractional representations. Assembling large sets of molecules represented in this way will result in sparse data, where a large portion of the set is null values. This consumes an excessive amount of computer memory which inhibits the size of data sets that can be used when constructing predictive models.In this study, we suggest a set of criteria for evaluation of random forest implementations to be used for in silico predictive modeling on sparse data sets, with regard to computer memory usage, model construction time and predictive accuracy.A novel random forest system was implemented to meet the suggested criteria, and experiments were made to compare our implementation to existing machine learning algorithms to establish our implementation‟s correctness. Experimental results show that our random forest implementation can create accurate prediction models on sparse datasets, with lower memory usage overhead than implementations using a common matrix representation, and in less time than existing random forest implementations evaluated against. We highlight design choices made to accommodate for sparse data structures and data sets in the random forest ensemble technique, and therein present potential improvements to feature selection in sparse data sets. / Program: Systemarkitekturutbildningen

A aquisição e o desempenho de linguagem humana é um processo pelo qual todas as pessoas passam. No entanto, esse processo não é completamente entendido, o que gera amplo espaço para pesquisa nessa área. Além disso, mesmo após o processo de aquisição da linguagem pela criança estar completo, ainda não há garantia de domínio da língua em suas diferentes modalidades, especialmente de leitura e escrita. Recentemente, em 2016, divulgou-se que 49,3% dos estudantes brasileiros não possuem proficiência de compreensão de leitura plena em português. Isso é particularmente importante ao considerarmos a quantidade de textos disponíveis, mas não acessíveis a pessoas com diferentes tipos de problemas de proficiência na língua. Sob o ponto de vista computacional, há estudos que visam modelar os processos de aquisição da linguagem e medir o nível do falante, leitor ou redator. Em vista disso, neste trabalho propomos uma abordagem computacional independente de idioma para modelar o nível de desenvolvimento linguístico de diferentes tipos de usuários da língua, de crianças e adultos, sendo a nossa proposta fortemente baseada em características linguísticas. Essas características são dependentes de corpora orais transcritos, no segmento de crianças, e de corpora escritos, no segmento de adultos. Para alcançar esse modelo abrangente, são considerados como objetivos a identificação de atributos e valores que diferenciam os níveis de desenvolvimento da linguagem do indivíduo, assim como o desenvolvimento de um modelo capaz de indicá-los. Para a identificação dos atributos, utilizamos métodos baseados em estatística, como o teste de hipóteses e divergência de distribuição. A fim de comprovar a abrangência da abordagem, realizamos experimentos com os corpora que espelham diferentes etapas do desenvolvimento da linguagem humana: (1) etapa de aquisição da linguagem oral de pela criança e (2) etapa pós aquisição, através da percepção de complexidade da linguagem escrita. Como resultados, obtivemos um grande conjunto anotado de dados sobre aquisição e desempenho de linguagem que podem contribuir para outros estudos. Assim como um perfil de atributos para os vários níveis de desenvolvimento. Também destacamos como resultados, os modelos computacionais que identificam textos quanto ao nível de desenvolvimento de linguagem. Em especial, o são resultados do trabalho o modelo de identificação de palavras complexas, que ultrapassou o estado da arte para o corpus estudado, e o modelo de identificação de idade de crianças que ultrapassou os baselines utilizados, incluindo uma medida clássica de desenvolvimento linguístico. / Language acquisition and language performance is a process by which all the people experience. However, this process is not completely understood, which creates room for research in this area. Moreover, even after the acquisition process by a child is completed, there is still no guarantee of language proficiency in different modalities, specially reading and writing. Recently, in 2016, OECD/PIAAC released that 49,3% of Brazilian students do not have written and read proficiency in Portuguese. This is more important when we take into account the large number of available text, but they are not accessible by people with different types of language proficiency issues. In computational point of view, there are some studies which aim to model the language acquisition process and measure the speaker level. For that, we propose an computational approach independent of language to model language development level of different types of language users, children and adults. In that sense our proposal is highly based on linguistics features. Those features dependents of transcript oral corpora from children and adults. To achieve this model, we considered aim to identify attributes and values able to differentiate between leves of development by an individual, as well the desenvolvimento of a model able to indicate them. The attribute identification are based on statistical methods such as hypothesis testing and divergence distribution. Aiming to validate our approach, we performed experiments with the corpora that reflect at different stages of development of human language: (1) oral language acquisition by a child and (2) post-acquisition stage, through the perception of difficulty of written language. With this work, we obtained a large corpus of annotated language acquisition data that can contribute to the acquisition of other studies. We also build an attribute profile of the development levels. From all of our results we highlight the computer models that identify texts and language development level. In particular, the complex word identification model that exceeded the state of the art for the studied corpus, and the children age identifier model, who exceeded the baselines, including a classic measure of language development.

Características de texturas atuam como bons descritores de imagens e podem ser empregadas em diversos problemas, como classificação e segmentação. Porém, quando o número de características é muito elevado, o reconhecimento de padrões pode ser prejudicado. A seleção de características contribui para a solução desse problema, podendo ser empregada tanto para redução da dimensionalidade como também para descobrir quais as melhores características de texturas para o tipo de imagem analisada. O objetivo deste trabalho é avaliar métodos ótimos e subótimos de seleção de características em problemas que envolvem texturas de imagens. Os algoritmos de seleção avaliados foram o branch and bound, a busca exaustiva e o sequential oating forward selection (SFFS). As funções critério empregadas na seleção foram a distância de Jeffries-Matusita e a taxa de acerto do classificador de distância mínima (CDM). As características de texturas empregadas nos experimentos foram obtidas com estatísticas de primeira ordem, matrizes de co-ocorrência e filtros de Gabor. Os experimentos realizados foram a classificação de regiôes de uma foto aérea de plantação de eucalipto, a segmentação não-supervisionada de mosaicos de texturas de Brodatz e a segmentação supervisionada de imagens médicas (MRI do cérebro). O branch and bound é um algoritmo ótimo e mais efiiente do que a busca exaustiva na maioria dos casos. Porém, continua sendo um algoritmo lento. Este trabalho apresenta uma nova estratégia para o branch and bound, nomeada floresta, que melhorou significativamente a eficiência do algoritmo. A avaliação dos métodos de seleção de características mostrou que os melhores subconjuntos foram aqueles obtidos com o uso da taxa de acerto do CDM. A busca exaustiva e o branch and bound, mesmo com a estratégia floresta, foram considerados inviáveis devido ao alto tempo de processamento nos casos em que o número de característica é muito grande. O SFFS apresentou os melhores resultados, pois, além de mais rápido, encontrou as soluções ótimas ou próximas das ótimas. Pôde-se concluir também que a precisão no reconhecimento de padrões aumenta com a redução do número de características e que os melhores subconjuntos freqüentemente são formados por características de texturas obtidas com técnicas diferentes / Texture features are eficient image descriptors and can be employed in a wide range of applications, such as classification and segmentation. However, when the number of features is considerably high, pattern recognition tasks may be compromised. Feature selection helps prevent this problem, as it can be used to reduce data dimensionality and reveal features which best characterise images under investigation. This work aims to evaluate optimal and suboptimal feature selection algorithms in the context of textural features extracted from images. Branch and bound, exhaustive search and sequential floating forward selection (SFFS) were the algorithms investigated. The criterion functions employed during selection were the Jeffries-Matusita (JM) distance and the minimum distance classifier (MDC) accuracy rate. Texture features were computed from first-order statistics, co-occurrence matrices and Gabor filters. Three different experiments have been conducted: classification of aerial picture of eucalyptus plantations, unsupervised segmentation of mosaics of Brodatz texture samples and supervised segmentation of MRI images of the brain. The branch and bound is an optimal algorithm and many times more eficient than exhaustive search. But is still time consuming. This work proposed a novel strategy for the branch and bound algorithm, named forest, which has considerably improved its performance. The evaluation of the feature selection methods has revealed that the best feature subsets were those computed by the MDC accuracy rate criterion function. Exhaustive search and branch and bound approaches have been considered unfeasible, due to their high processing times, especially for high dimensional data. This statement holds even for the branch and bound with the forest strategy. The SFFS approach yielded the best results. Not only was it faster, as it also was capable of finding the optimal or nearly optimal solutions. Finally, it has been observed that the precision of pattern recognition tasks increases as the number of features decreases and that the best feature subsets are those which possess features computed from distinct texture feature methods

O reconhecimento de faces é uma área de pesquisa desafiadora que abre portas para a implementação de aplicações muito promissoras. Embora muitos algoritmos eficientes e robustos já tenham sido propostos, ainda restam vários desafios. Dentre os principais obstáculos a serem uperados, está a obtenção de uma representação robusta e compacta de faces que possibilite distinguir os indivíduos rapidamente. Visando abordar esse problema, foi realizado um estudo de técnicas de reconhecimento estatístico de padrões, principalmente na área de redução de dimensionalidade dos dados, além de uma revisão de métodos de reconhecimento de faces. Foi proposto (em colaboração com a pesquisadora Isabelle Bloch) um método de seleção de características que une um algoritmo de busca eficiente (métodos de busca seqüencial flutuante) com uma medida de distância entre conjuntos nebulosos (distância nebulosa baseada em tolerância). Essa medida de distância possui diversas vantagens, sendo possível considerar as diferentes tipicalidades de cada padrão dos conjuntos de modo a permitir a obtenção de bons resultados mesmo com conjuntos com sobreposição. Os resultados preliminares com dados sintéticos mostraram o caráter promissor dessa abordagem. Com o objetivo de verificar a eficiência de tal técnica com dados reais, foram efetuados testes com reconhecimento de pessoas usando imagens da região dos olhos. Nesse caso, em se tratando de um problema com mais de duas classes, nós propusemos uma nova função critério inspirada na distância supracitada. Além disso foi proposto (juntamente com o estudante de mestrado Rogério S. Feris) um esquema de reconhecimento a partir de seqüências de vídeo. Esse esquema inclui a utilização de um método eficiente de rastreamento de características faciais (Gabor Wavelet Networks) e o método proposto anteriormente para seleção de características. Dentro desse contexto, o trabalho desenvolvido nesta dissertação implementa uma parte dos módulos desse esquema. / Face recognition is an instigating research field that may lead to the development of many promising applications. Although many efficient and robust algorithms have been developed in this area, there are still many challenges to be overcome. In particular, a robust and compact face representation is still to be found, which would allow for quick classification of different individuals. In order to address this problem, we first studied pattern recognition techniques, especially regarding dimensionality reduction, followed by the main face recognition methods. We introduced a new feature selection approach in collaboration with the researcher Isabelle Bloch (TSI-ENST-Paris), that associates an efficient searching algorithm (sequential floating search methods), with a tolerance-based fuzzy distance. This distance measure presents some nice features for dealing with the tipicalities of each pattern in the sets, so that good results can be attained even when the sets are overlapping. Preliminary results with synthetic data have demonstrated that this method is quite promising. In order to verify the efficiency of this technique with real data, we applied it for improving the performance of a person recognition system based on eye images. Since this problem involves more than two classes, we also developed a new criterion function based on the above-mentioned distance. Moreover, we proposed (together with Rogério S. Feris) a system for person recognition based on video sequences. This mechanism includes the development of an efficient method for facial features tracking, in addition to our method for feature selection. In this context, the work presented here constitutes part of the proposed system.

Along with the improvement of data acquisition techniques and the increasing computational capacity of computers, the dimensionality of the data grows higher. Pattern recognition methods have to deal with samples consisting of thousands of features and the reduction of their dimensionality becomes crucial to make them tractable. Feature selection is a technique for removing the irrelevant and noisy features and selecting a subset of features which describe better the samples and produce a better classification performance. It is becoming an essential part of most pattern recognition applications. / In this thesis we propose a feature selection method for supervised classification. The main contribution is the efficient use of information theory, which provides a solid theoretical framework for measuring the relation between the classes and the features. Mutual information is considered to be the best measure for such purpose. Traditionally it has been measured for ranking single features without taking into account the entire set of selected features. This is due to the computational complexity involved in estimating the mutual information. However, in most data sets the features are not independent and their combination provides much more information about the class, than the sum of their individual prediction power. / Methods based on density estimation can only be used for data sets with a very high number of samples and low number of features. Due to the curse of dimensionality, in a multi-dimensional feature space the amount of samples required for a reliable density estimation is very high. For this reason we analyse the use of different estimation methods which bypass the density estimation and estimate entropy directly from the set of samples. These methods allow us to efficiently evaluate sets of thousands of features. / For high-dimensional feature sets another problem is the search order of the feature space. All non-prohibitive computational cost algorithms search for a sub-optimal feature set. Greedy algorithms are the fastest and are the ones which incur less overfitting. We show that from the information theoretical perspective, a greedy backward selection algorithm conserves the amount of mutual information, even though the feature set is not the minimal one. / We also validate our method in several real-world applications. We apply feature selection to omnidirectional image classification through a novel approach. It is appearance-based and we select features from a bank of filters applied to different parts of the image. The context of the task is place recognition for mobile robotics. Another set of experiments are performed on microarrays from gene expression databases. The classification problem aims to predict the disease of a new patient. We present a comparison of the classification performance and the algorithms we present showed to outperform the existing ones. Finally, we succesfully apply feature selection to spectral graph classification. All the features we use are for unattributed graphs, which constitutes a contribution to the field. We also draw interesting conclusions about which spectral features matter most, under different experimental conditions. In the context of graph classification we also show important is the precise estimation of mutual information and we analyse its impact on the final classification results.

This study is an investigation if a recurrent long short-term memory (LSTM) based neural network can be used to estimate the battery capacity in electrical cars. There is an enormous interest in finding the underlying reasons why and how Lithium-ion batteries ages and this study is a part of this broader question. The research questions that have been answered are how well a LSTM model estimates the battery capacity, how the LSTM model is performing compared to a linear model and what parameters that are important when estimating the capacity. There have been other studies covering similar topics but only a few that has been performed on a real data set from real cars driving. With a data science approach, it was discovered that the LSTM model indeed is a powerful model to use for estimation the capacity. It had better accuracy than a linear regression model, but the linear regression model still gave good results. The parameters that implied to be important when estimating the capacity were logically related to the properties of a Lithium-ion battery.En studie över hur väl ett återkommande neuralt nätverk kan estimera kapaciteten hos Litium-ion batteri hos elektroniska fordon, när en en datavetenskaplig strategi har använts.

This thesis proposes a series of multi-label learning algorithms for classification and feature selection implemented on the Apache Spark distributed computing model. Five approaches for determining the optimal architecture to speed up multi-label learning methods are presented. These approaches range from local parallelization using threads to distributed computing using independent or shared memory spaces. It is shown that the optimal approach performs hundreds of times faster than the baseline method. Three distributed multi-label k nearest neighbors methods built on top of the Spark architecture are proposed: an exact iterative method that computes pair-wise distances, an approximate tree-based method that indexes the instances across multiple nodes, and an approximate local sensitive hashing method that builds multiple hash tables to index the data. The results indicated that the predictions of the tree-based method are on par with those of an exact method while reducing the execution times in all the scenarios. The aforementioned method is then used to evaluate the quality of a selected feature subset. The optimal adaptation for a multi-label feature selection criterion is discussed and two distributed feature selection methods for multi-label problems are proposed: a method that selects the feature subset that maximizes the Euclidean norm of individual information measures, and a method that selects the subset of features maximizing the geometric mean. The results indicate that each method excels in different scenarios depending on type of features and the number of labels. Rigorous experimental studies and statistical analyses over many multi-label metrics and datasets confirm that the proposals achieve better performances and provide better scalability to bigger data than the methods compared in the state of the art.

Cette thèse de doctorat est consacrée à un sujet de recherche très porteur : la Catégorisation générique d’objets Visuels (VOC). En effet, les applications possibles sont très nombreuses, incluant l’indexation d’images et de vidéos, la vidéo surveillance, le contrôle d’accès de sécurité, le soutien à la conduite automobile, etc. En raison de ses nombreux verrous scientifiques, ce sujet est encore considéré comme l’un des problèmes les plus difficiles en vision par ordinateur et en reconnaissance de formes. Dans ce contexte, nous avons proposé dans ce travail de thèse plusieurs contributions, en particulier concernant les deux principaux éléments des méthodes résolvant les problèmes de VOC, notamment la sélection des descripteurs et la représentation d’images. Premièrement, un algorithme nomme "Embedded Sequential Forward feature Selection"(ESFS) a été proposé pour VOC. Son but est de sélectionner les descripteurs les plus discriminants afin d’obtenir une bonne performance pour la catégorisation. Il est principalement basé sur la méthode de recherche sous-optimale couramment utilisée "Sequential Forward Selection" (SFS), qui repose sur le principe simple d’ajouter progressivement les descripteurs les plus pertinents. Cependant, ESFS non seulement ajoute progressivement les descripteurs les plus pertinents à chaque étape mais de plus les fusionne d’une manière intégrée grâce à la notion de fonctions de masses combinées empruntée à la théorie de l’évidence qui offre également l’avantage d’obtenir un coût de calcul beaucoup plus faible que celui de SFS original. Deuxièmement, nous avons proposé deux nouvelles représentations d’images pour modéliser le contenu visuel d’une image : la Représentation d’Image basée sur la Modélisation Polynomiale et les Mesures Statistiques, appelées respectivement PMIR et SMIR. Elles permettent de surmonter l’inconvénient principal de la méthode populaire "bag of features" qui est la difficulté de fixer la taille optimale du vocabulaire visuel. Elles ont été testées avec nos descripteurs bases région ainsi que les descripteurs SIFT. Deux stratégies différentes de fusion, précoce et tardive, ont également été considérées afin de fusionner les informations venant des "canaux «différents représentés par les différents types de descripteurs. Troisièmement, nous avons proposé deux approches pour VOC en s’appuyant sur la représentation sparse. La première méthode est reconstructive (R_SROC) alors que la deuxième est reconstructive et discriminative (RD_SROC). En effet, le modèle de représentation sparse a été utilisé originalement dans le domaine du traitement du signal comme un outil puissant pour acquérir, représenter et compresser des signaux de grande dimension. Ainsi, nous avons proposé une adaptation de ces principes intéressants au problème de VOC. R_SROC repose sur l’hypothèse intuitive que l’image peut être représentée par une combinaison linéaire des images d’apprentissage de la même catégorie. [...] / This thesis is dedicated to the active research topic of generic Visual Object Categorization(VOC), which can be widely used in many applications such as videoindexation and retrieval, video monitoring, security access control, automobile drivingsupport etc. Due to many realistic difficulties, it is still considered to be one ofthe most challenging problems in computer vision and pattern recognition. In thiscontext, we have proposed in this thesis our contributions, especially concerning thetwo main components of the methods addressing VOC problems, namely featureselection and image representation.Firstly, an Embedded Sequential Forward feature Selection algorithm (ESFS)has been proposed for VOC. Its aim is to select the most discriminant features forobtaining a good performance for the categorization. It is mainly based on thecommonly used sub-optimal search method Sequential Forward Selection (SFS),which relies on the simple principle to add incrementally most relevant features.However, ESFS not only adds incrementally most relevant features in each stepbut also merges them in an embedded way thanks to the concept of combinedmass functions from the evidence theory which also offers the benefit of obtaining acomputational cost much lower than the one of original SFS.Secondly, we have proposed novel image representations to model the visualcontent of an image, namely Polynomial Modeling and Statistical Measures basedImage Representation, called PMIR and SMIR respectively. They allow to overcomethe main drawback of the popular "bag of features" method which is the difficultyto fix the optimal size of the visual vocabulary. They have been tested along withour proposed region based features and SIFT. Two different fusion strategies, earlyand late, have also been considered to merge information from different "channels"represented by the different types of features.Thirdly, we have proposed two approaches for VOC relying on sparse representation,including a reconstructive method (R_SROC) as well as a reconstructiveand discriminative one (RD_SROC). Indeed, sparse representation model has beenoriginally used in signal processing as a powerful tool for acquiring, representingand compressing the high-dimensional signals. Thus, we have proposed to adaptthese interesting principles to the VOC problem. R_SROC relies on the intuitiveassumption that an image can be represented by a linear combination of trainingimages from the same category. Therefore, the sparse representations of images arefirst computed through solving the ℓ1 norm minimization problem and then usedas new feature vectors for images to be classified by traditional classifiers such asSVM. To improve the discrimination ability of the sparse representation to betterfit the classification problem, we have also proposed RD_SROC which includes adiscrimination term, such as Fisher discrimination measure or the output of a SVMclassifier, to the standard sparse representation objective function in order to learna reconstructive and discriminative dictionary. Moreover, we have also proposedChapter 0. Abstractto combine the reconstructive and discriminative dictionary and the adapted purereconstructive dictionary for a given category so that the discrimination power canfurther be increased.The efficiency of all the methods proposed in this thesis has been evaluated onpopular image datasets including SIMPLIcity, Caltech101 and Pascal2007.