Au cours de ces dernières années, les matrices de covariance ont montré leur intérêt dans de nombreuses applications en traitement du signal et de l'image.Les travaux présentés dans cette thèse se concentrent sur l'utilisation de ces matrices comme descripteurs pour la classification. Dans ce contexte, des algorithmes robustes de classification sont proposés en développant les aspects suivants.Tout d'abord, des estimateurs robustes de la matrice de covariance sont utilisés afin de réduire l'impact des observations aberrantes. Puis, les distributions Riemannienne Gaussienne et de Laplace, ainsi que leur extension au cas des modèles de mélange, sont considérés pour la modélisation des matrices de covariance.Les algorithmes de type k-moyennes et d'espérance-maximisation sont étendus au cas Riemannien pour l'estimation de paramètres de ces lois : poids, centroïdes et paramètres de dispersion. De plus, un nouvel estimateur du centroïde est proposé en s'appuyant sur la théorie des M-estimateurs : l'estimateur de Huber. En outre,des descripteurs appelés vecteurs Riemannien de Fisher sont introduits afin de modéliser les images non-stationnaires. Enfin, un test d'hypothèse basé sur la distance géodésique est introduit pour réguler la probabilité de fausse alarme du classifieur.Toutes ces contributions sont validées en classification d'images de texture, de signaux du cerveau, et d'images polarimétriques radar simulées et réelles. / In the recent years, covariance matrices have demonstrated their interestin a wide variety of applications in signal and image processing. The workpresented in this thesis focuses on the use of covariance matrices as signatures forrobust classification. In this context, a robust classification workflow is proposed,resulting in the following contributions.First, robust covariance matrix estimators are used to reduce the impact of outlierobservations, during the estimation process. Second, the Riemannian Gaussianand Laplace distributions as well as their mixture model are considered to representthe observed covariance matrices. The k-means and expectation maximization algorithmsare then extended to the Riemannian case to estimate their parameters, thatare the mixture's weight, the central covariance matrix and the dispersion. Next,a new centroid estimator, called the Huber's centroid, is introduced based on thetheory of M-estimators. Further on, a new local descriptor named the RiemannianFisher vector is introduced to model non-stationary images. Moreover, a statisticalhypothesis test is introduced based on the geodesic distance to regulate the classification false alarm rate. In the end, the proposed methods are evaluated in thecontext of texture image classification, brain decoding, simulated and real PolSARimage classification.

No description available.

The central theme of the thesis is to study linear and non linear SVM formulations in the presence of uncertain observations. The main contribution of this thesis is to derive robust classfiers from partial knowledge of the underlying uncertainty. In the case of linear classification, a new bounding scheme based on Bernstein inequality has been proposed, which models interval-valued uncertainty in a less conservative fashion and hence is expected to generalize better than the existing methods. Next, potential of partial information such as bounds on second order moments along with support information has been explored. Bounds on second order moments make the resulting classifiers robust to moment estimation errors. Uncertainty in the dataset will lead to uncertainty in the kernel matrices. A novel distribution free large deviation inequality has been proposed which handles uncertainty in kernels through co-positive programming in a chance constraint setting. Although such formulations are NP hard, under several cases of interest the problem reduces to a convex program. However, the independence assumption mentioned above, is restrictive and may not always define a valid uncertain kernel. To alleviate this problem an affine set based alternative is proposed and using a robust optimization framework the resultant problem is posed as a minimax problem. In both the cases of Chance Constraint Program or Robust Optimization (for non-linear SVM), mirror descent algorithm (MDA) like procedures have been applied.