L’imagerie hyperspectrale (HS) consiste à acquérir une même scène dans plusieurs centaines de bandes spectrales contiguës (dimensions d'un cube de données), ce qui a conduit à trois types d'applications pertinentes, telles que la détection de cibles, la classification et le démélange spectral. Cependant, tandis que les capteurs hyperspectraux fournissent une information spectrale abondante, leur résolution spatiale est généralement plus limitée. Ainsi, la fusion d’une image HS avec d'autres images à haute résolution de la même scène, telles que les images multispectrales (MS) ou panchromatiques (PAN) est un problème intéressant. Le problème de fusionner une image HS de haute résolution spectrale mais de résolution spatiale limitée avec une image auxiliaire de haute résolution spatiale mais de résolution spectrale plus limitée (parfois qualifiée de fusion multi-résolution) a été exploré depuis de nombreuses années. D'un point de vue applicatif, ce problème est également important et est motivé par ceratins projets, comme par exemple le project Japonais HISIU, qui vise à fusionner des images MS et HS recalées acquises pour la même scène avec les mêmes conditions. Les techniques de fusion bayésienne permettent une interprétation intuitive du processus de fusion via la définition de la loi a posteriori de l’image à estimer (qui est de hautes résolutions spatiale et spectrale). Puisque le problème de fusion est généralement mal posé, l’inférence bayésienne offre un moyen pratique pour régulariser le problème en définissant une loi a priori adaptée à la scène d'intérêt. Les différents chapitres de cette thèse sont résumés ci-dessous. Le introduction présente le modèle général de fusion et les hypothèses statistiques utilisées pour les images multi-bandes observées, c’est-à-dire les images HS, MS ou PAN. Les images observées sont des versions dégradées de l'image de référence (à hautes résolutions spatiale et spectrale) qui résultent par exemple d’un flou spatial et spectral et/ou d’un sous-échantillonnage liés aux caractéristiques des capteurs. Les propriétés statistiques des mesures sont alors obtenues directement à partir d’un modèle linéaire traduisant ces dégradations et des propriétés statistiques du bruit. Le chapitre 1 s’intéresse à une technique de fusion bayésienne pour les images multi-bandes de télédétection, à savoir pour les images HS, MS et PAN. Tout d'abord, le problème de fusion est formulé dans un cadre d'estimation bayésienne. Une loi a priori Gaussienne exploitant la géométrie du problème est définie et un algorithme d’estimation Bayésienne permettant d’estimer l’image de référence est étudié. Pour obtenir des estimateurs Bayésiens liés à la distribution postérieure résultant, deux algorithmes basés sur échantillonnage de Monte Carlo et l'optimisation stratégie ont été développés. Le chapitre 2 propose une approche variationnelle pour la fusion d’images HS et MS. Le problème de fusion est formulé comme un problème inverse dont la solution est l'image d’intérêt qui est supposée vivre dans un espace de dimension résuite. Un terme de régularisation imposant des contraintes de parcimonie est défini avec soin. Ce terme traduit le fait que les patches de l'image cible sont bien représentés par une combinaison linéaire d’atomes appartenant à un dictionnaire approprié. Les atomes de ce dictionnaire et le support des coefficients des décompositions des patches sur ces atomes sont appris à l’aide de l’image de haute résolution spatiale. Puis, conditionnellement à ces dictionnaires et à ces supports, le problème de fusion est résolu à l’aide d’un algorithme d’optimisation alternée (utilisant l’algorithme ADMM) qui estime de manière itérative l’image d’intérêt et les coefficients de décomposition. / Hyperspectral (HS) imaging, which consists of acquiring a same scene in several hundreds of contiguous spectral bands (a three dimensional data cube), has opened a new range of relevant applications, such as target detection [MS02], classification [C.-03] and spectral unmixing [BDPD+12]. However, while HS sensors provide abundant spectral information, their spatial resolution is generally more limited. Thus, fusing the HS image with other highly resolved images of the same scene, such as multispectral (MS) or panchromatic (PAN) images is an interesting problem. The problem of fusing a high spectral and low spatial resolution image with an auxiliary image of higher spatial but lower spectral resolution, also known as multi-resolution image fusion, has been explored for many years [AMV+11]. From an application point of view, this problem is also important as motivated by recent national programs, e.g., the Japanese next-generation space-borne hyperspectral image suite (HISUI), which fuses co-registered MS and HS images acquired over the same scene under the same conditions [YI13]. Bayesian fusion allows for an intuitive interpretation of the fusion process via the posterior distribution. Since the fusion problem is usually ill-posed, the Bayesian methodology offers a convenient way to regularize the problem by defining appropriate prior distribution for the scene of interest. The aim of this thesis is to study new multi-band image fusion algorithms to enhance the resolution of hyperspectral image. In the first chapter, a hierarchical Bayesian framework is proposed for multi-band image fusion by incorporating forward model, statistical assumptions and Gaussian prior for the target image to be restored. To derive Bayesian estimators associated with the resulting posterior distribution, two algorithms based on Monte Carlo sampling and optimization strategy have been developed. In the second chapter, a sparse regularization using dictionaries learned from the observed images is introduced as an alternative of the naive Gaussian prior proposed in Chapter 1. instead of Gaussian prior is introduced to regularize the ill-posed problem. Identifying the supports jointly with the dictionaries circumvented the difficulty inherent to sparse coding. To minimize the target function, an alternate optimization algorithm has been designed, which accelerates the fusion process magnificently comparing with the simulation-based method. In the third chapter, by exploiting intrinsic properties of the blurring and downsampling matrices, a much more efficient fusion method is proposed thanks to a closed-form solution for the Sylvester matrix equation associated with maximizing the likelihood. The proposed solution can be embedded into an alternating direction method of multipliers or a block coordinate descent method to incorporate different priors or hyper-priors for the fusion problem, allowing for Bayesian estimators. In the last chapter, a joint multi-band image fusion and unmixing scheme is proposed by combining the well admitted linear spectral mixture model and the forward model. The joint fusion and unmixing problem is solved in an alternating optimization framework, mainly consisting of solving a Sylvester equation and projecting onto a simplex resulting from the non-negativity and sum-to-one constraints. The simulation results conducted on synthetic and semi-synthetic images illustrate the advantages of the developed Bayesian estimators, both qualitatively and quantitatively.

In this dissertation, we discuss the problem of robust linear subspace estimation using low-rank optimization and propose three formulations of it. We demonstrate how these formulations can be used to solve fundamental computer vision problems, and provide superior performance in terms of accuracy and running time. Consider a set of observations extracted from images (such as pixel gray values, local features, trajectories . . . etc). If the assumption that these observations are drawn from a liner subspace (or can be linearly approximated) is valid, then the goal is to represent each observation as a linear combination of a compact basis, while maintaining a minimal reconstruction error. One of the earliest, yet most popular, approaches to achieve that is Principal Component Analysis (PCA). However, PCA can only handle Gaussian noise, and thus suffers when the observations are contaminated with gross and sparse outliers. To this end, in this dissertation, we focus on estimating the subspace robustly using low-rank optimization, where the sparse outliers are detected and separated through the `1 norm. The robust estimation has a two-fold advantage: First, the obtained basis better represents the actual subspace because it does not include contributions from the outliers. Second, the detected outliers are often of a specific interest in many applications, as we will show throughout this thesis. We demonstrate four different formulations and applications for low-rank optimization. First, we consider the problem of reconstructing an underwater sequence by removing the iii turbulence caused by the water waves. The main drawback of most previous attempts to tackle this problem is that they heavily depend on modelling the waves, which in fact is ill-posed since the actual behavior of the waves along with the imaging process are complicated and include several noise components; therefore, their results are not satisfactory. In contrast, we propose a novel approach which outperforms the state-of-the-art. The intuition behind our method is that in a sequence where the water is static, the frames would be linearly correlated. Therefore, in the presence of water waves, we may consider the frames as noisy observations drawn from a the subspace of linearly correlated frames. However, the noise introduced by the water waves is not sparse, and thus cannot directly be detected using low-rank optimization. Therefore, we propose a data-driven two-stage approach, where the first stage “sparsifies” the noise, and the second stage detects it. The first stage leverages the temporal mean of the sequence to overcome the structured turbulence of the waves through an iterative registration algorithm. The result of the first stage is a high quality mean and a better structured sequence; however, the sequence still contains unstructured sparse noise. Thus, we employ a second stage at which we extract the sparse errors from the sequence through rank minimization. Our method converges faster, and drastically outperforms state of the art on all testing sequences. Secondly, we consider a closely related situation where an independently moving object is also present in the turbulent video. More precisely, we consider video sequences acquired in a desert battlefields, where atmospheric turbulence is typically present, in addition to independently moving targets. Typical approaches for turbulence mitigation follow averaging or de-warping techniques. Although these methods can reduce the turbulence, they distort the independently moving objects which can often be of great interest. Therefore, we address the iv problem of simultaneous turbulence mitigation and moving object detection. We propose a novel three-term low-rank matrix decomposition approach in which we decompose the turbulence sequence into three components: the background, the turbulence, and the object. We simplify this extremely difficult problem into a minimization of nuclear norm, Frobenius norm, and `1 norm. Our method is based on two observations: First, the turbulence causes dense and Gaussian noise, and therefore can be captured by Frobenius norm, while the moving objects are sparse and thus can be captured by `1 norm. Second, since the object’s motion is linear and intrinsically different than the Gaussian-like turbulence, a Gaussian-based turbulence model can be employed to enforce an additional constraint on the search space of the minimization. We demonstrate the robustness of our approach on challenging sequences which are significantly distorted with atmospheric turbulence and include extremely tiny moving objects. In addition to robustly detecting the subspace of the frames of a sequence, we consider using trajectories as observations in the low-rank optimization framework. In particular, in videos acquired by moving cameras, we track all the pixels in the video and use that to estimate the camera motion subspace. This is particularly useful in activity recognition, which typically requires standard preprocessing steps such as motion compensation, moving object detection, and object tracking. The errors from the motion compensation step propagate to the object detection stage, resulting in miss-detections, which further complicates the tracking stage, resulting in cluttered and incorrect tracks. In contrast, we propose a novel approach which does not follow the standard steps, and accordingly avoids the aforementioned diffi- culties. Our approach is based on Lagrangian particle trajectories which are a set of dense trajectories obtained by advecting optical flow over time, thus capturing the ensemble motions v of a scene. This is done in frames of unaligned video, and no object detection is required. In order to handle the moving camera, we decompose the trajectories into their camera-induced and object-induced components. Having obtained the relevant object motion trajectories, we compute a compact set of chaotic invariant features, which captures the characteristics of the trajectories. Consequently, a SVM is employed to learn and recognize the human actions using the computed motion features. We performed intensive experiments on multiple benchmark datasets, and obtained promising results. Finally, we consider a more challenging problem referred to as complex event recognition, where the activities of interest are complex and unconstrained. This problem typically pose significant challenges because it involves videos of highly variable content, noise, length, frame size . . . etc. In this extremely challenging task, high-level features have recently shown a promising direction as in [53, 129], where core low-level events referred to as concepts are annotated and modelled using a portion of the training data, then each event is described using its content of these concepts. However, because of the complex nature of the videos, both the concept models and the corresponding high-level features are significantly noisy. In order to address this problem, we propose a novel low-rank formulation, which combines the precisely annotated videos used to train the concepts, with the rich high-level features. Our approach finds a new representation for each event, which is not only low-rank, but also constrained to adhere to the concept annotation, thus suppressing the noise, and maintaining a consistent occurrence of the concepts in each event. Extensive experiments on large scale real world dataset TRECVID Multimedia Event Detection 2011 and 2012 demonstrate that our approach consistently improves the discriminativity of the high-level features by a significant margin.