Cette thèse apporte des solutions théoriques et pratiques à deux problèmes soulevés par la photographie numérique en présence de mouvement, et par la photographie infrarouge. La photographie d'objets en mouvement semblait ne pouvoir se faire qu'avec des temps d'exposition très courts, jusqu'à ce que deux travaux révolutionnaires proposent deux nouveaux types de caméra permettant un temps d'exposition arbitraire. Le flutter shutter de Agrawal et al. crée en effet un flou inversible, grâce à un obturateur aux séquences d'ouverture-fermeture bie{\it n choisies. Le motion invariant photography de Levin et al. obtient ce même effet avec une accélération constante de la caméra. Les deux méthodes suivent ainsi un nouveau paradigme, la computational photography, selon lequel les caméras sont repensées, car elles incluent un traitement numérique sophistiqué. Cette thèse propose une méthode pour évaluer la qualité image des nouvelles caméras. Le fil conducteur de l'analyse est donc l'évaluation du SNR (signal to noise ratio) de l'image obtenue après déconvolution. La théorie fournit des formules explicites pour le SNR, soulève deux paradoxes de ces caméras, et les résout. Elle permet d'obtenir le modèle de mouvement sous-jacent à chaque flutter shutter, notamment tous ceux qui sont brevetés. Une seconde partie plus brève aborde le problème de qualité principal en imagerie vidéo infrarouge, la non-uniformité. Il s'agit d'un bruit évolutif et structuré en colonnes causé par le capteur. La conclusion des travaux est qu'il est non seulement possible mais également efficace et robuste d'effectuer la correction sur une seule image. Cela permet de contourner le problème récurrent des "ghost artifacts"résultant d'une incohérence du traitement par rapport au modèle d'acquisition. / This thesis provides theoretical and practical solutions to two problems raised by digital photography of moving scenes, and infrared photography. Until recently photographing moving objects could only be done using short exposure times. Yet, two recent groundbreaking works have proposed two new designs of camera allowing arbitrary exposure times. The flutter shutter of Agrawal et al. creates an invertible motion blur by using a clever shutter technique to interrupt the photon flux during the exposure time according to a well chosen binary sequence. The motion-invariant photography of Levin et al. gets the same result by accelerating the camera at a constant rate. Both methods follow computational photography as a new paradigm. The conception of cameras is rethought to include sophisticated digital processing. This thesis proposes a method for evaluating the image quality of these new cameras. The leitmotiv of the analysis is the SNR (signal to noise ratio) of the image after deconvolution. It gives the efficiency of these new camera design in terms of image quality. The theory provides explicit formulas for the SNR. It raises two paradoxes of these cameras, and resolves them. It provides the underlying motion model of each flutter shutter, including patented ones. A shorter second part addresses the the main quality problem in infrared video imaging, the non-uniformity. This perturbation is a time-dependent noise caused by the infrared sensor, structured in columns. The conclusion of this work is that it is not only possible but also efficient and robust to perform the correction on a single image. This permits to ensure the absence of ``ghost artifacts'', a classic of the literature on the subject, coming from inadequate processing relative to the acquisition model.

This thesis provides theoretical and practical solutions to two problems raised by digital photography of moving scenes, and infrared photography. Until recently photographing moving objects could only be done using short exposure times. Yet, two recent groundbreaking works have proposed two new designs of camera allowing arbitrary exposure times. The flutter shutter of Agrawal et al. creates an invertible motion blur by using a clever shutter technique to interrupt the photon flux during the exposure time according to a well chosen binary sequence. The motion-invariant photography of Levin et al. gets the same result by accelerating the camera at a constant rate. Both methods follow computational photography as a new paradigm. The conception of cameras is rethought to include sophisticated digital processing. This thesis proposes a method for evaluating the image quality of these new cameras. The leitmotiv of the analysis is the SNR (signal to noise ratio) of the image after deconvolution. It gives the efficiency of these new camera design in terms of image quality. The theory provides explicit formulas for the SNR. It raises two paradoxes of these cameras, and resolves them. It provides the underlying motion model of each flutter shutter, including patented ones. A shorter second part addresses the the main quality problem in infrared video imaging, the non-uniformity. This perturbation is a time-dependent noise caused by the infrared sensor, structured in columns. The conclusion of this work is that it is not only possible but also efficient and robust to perform the correction on a single image. This permits to ensure the absence of ''ghost artifacts'', a classic of the literature on the subject, coming from inadequate processing relative to the acquisition model.

In the past decade, camera manufacturers have offered smaller form factors, smaller pixel sizes (leading to higher resolution images), and faster processing chips to increase the performance of consumer cameras. However, these conventional approaches have failed to capitalize on the spatio-temporal redundancy inherent in images, nor have they adequately provided a solution for finding $3$D point correspondences for cameras sampling different bands of the visible spectrum. In this thesis, we pose the following question---given the repetitious nature of image patches, and appropriate camera architectures, can statistical models be used to increase temporal, structural, or spectral resolution? While many techniques have been suggested to tackle individual aspects of this question, the proposed solutions either require prohibitively expensive hardware modifications and/or require overly simplistic assumptions about the geometry of the scene. We propose a two-stage solution to facilitate image reconstruction; 1) design a linear camera system that optically encodes scene information and 2) recover full scene information using prior models learned from statistics of natural images. By leveraging the tendency of small regions to repeat throughout an image or video, we are able to learn prior models from patches pulled from exemplar images. The quality of this approach will be demonstrated for two application domains, using low-speed video cameras for high-speed video acquisition and multi-spectral fusion using an array of cameras. We also investigate a conventional approach for finding 3D correspondence that enables a generalized assorted array of cameras to operate in multiple modalities, including multi-spectral, high dynamic range, and polarization imaging of dynamic scenes.