Topics in Multiframe Superresolution Restoration

Multiframe superresolution restoration refers to methods for still image and video restoration from multiple low-resolution, degraded observed images derived from an underlying scene. These methods extend classical single frame image restoration methods by simultaneously utilizing information from multiple observed images to achieve restoration at resolutions higher than that of the original data. Since multiframe methods are capable of restoring spectral information beyond that which is available in any of the observed images, they are referred to as superresolution methods. Multiframe superresolution restoration methods register the observed images to a common reference frame in order to formulate multiple observed data constraints for a given region of the restoration. The image registration process thus requires knowledge of the visual motion occurring in the observed image sequence. Since this typically is not known, the motion information must be estimated from the observed image sequence in order to effect the restoration process. An exhaustive survey of the multiframe superresolution restoration literature is presented along with some of the necessary background material. A detailed introduction to the problem of motion estimation is also provided. The performance characteristics of a common class of block-matching subpixel motion estimators is examined in an effort to understand the limitations that these motion estimators impose on the performance of multiframe superresolution restoration frameworks. A new Bayesian multiframe superresolution restoration procedure, which computes multiple restored frames simultaneously, is introduced. The proposed method enables the incorporation of temporal as well as spatial constraints on the restored image sequence. A novel approach for incorporating general interframe motion and linear, but spatially-varying, degradations in an observation matrix relating the observed low-resolution image sequence to the unknown restoration is presented. Stemming from the work conducted on multiframe superresolution restoration, the equivalence between two mathematical models describing a pinhole camera used in the computer graphics and computer vision communities is proved.