| With the advent of frame grabbers capable of acquiring multiple video frames, a great deal of attention is being directed towards developing digital video processing algorithms. These algorithms can be applied to increase the resolution of video, and also to print high-definition still images from interlaced or low-resolution video. Tasks such as these present a multi-faceted problem, which generally necessitates a sampling lattice up-conversion, as well as blur and noise removal. Three components of this problem, namely standards conversion, image sequence restoration, and super-resolution image reconstruction, have received considerable, yet separate treatment in the literature. Standards conversion treats the conversion from one spatio-temporal sampling lattice to another. The main focus is to remove the effects of aliasing without sacrificing resolution. Deblurring, however, is generally not done. The goal of image sequence restoration is to remover blur and noise, but in this case, the progressive input and output sampling lattices are the same. Super-resolution reconstruction is basically the image sequence restoration problem, where the output sampling lattice is a higher density progressive lattice than that of the input. In this thesis, a unifying framework is presented in which all of these problems can be simultaneously addressed. Within this framework, a projections onto convex sets based algorithm is proposed to solve this very general problem. In specific cases where only a subset of the most general problem needs to be solved, optimized algorithms are proposed. For deinterlacing in the presence of a dominant motion, such as camera jitter, a robust solution is presented. The dominant motion is first compensated for in a global manner, and then a novel motion detector, which uses adaptive thresholding, is applied to identify and correct for regions not undergoing the dominant motion. For the problem of standards conversion in the presence of accelerated motion, analysis in the short-time spectral domain is carried out. A filter design strategy is then provided based on this analysis. For image sequence restoration, and simultaneous deinterlacing and restoration, a new spatio-temporal reduced-order model Kalman filter is proposed, which is applied to the case of linear shift-varying blurring. In contrast to the existing literature, both motion-compensation and reduced-order state modeling are achieved by augmenting the observation equation, as opposed to modifying the state-transition equation. |
