| Digital images and digital video are inherently imperfect. For a scene to be converted to a digital image, it must be discretized in two spatial dimensions and discretized in value; i.e., the image must have finite resolution in width and height, and its pixel values must have a finite number of levels. Dynamic scenes must be discretized in time prior to conversion to digital video. Thus, there is uncertainty in four dimensions: width, height, time, and value. Furthermore, lossy compression is often employed to reduce the amount of storage or bandwidth that images and video consume, which leads to further loss of information and the introduction of artifacts. This dissertation discusses methods of recovering information lost due to quantization and compression. In particular, a method of regaining information lost due to limitations of a camera's dynamic range is proposed, which shows how to alleviate the bit-depth limitations of typical cameras. A spatial-domain noise model is derived to describe the quantization error introduced by compression, and is used in probabilistic formulations that address the spatial and temporal limitations of compressed images and video, in particular spatial resolution, temporal resolution, and spatial compression artifacts. The underlying focus of the work is the recovery of high-quality image and video data from observations that are limited in quality by the capture process itself, limited in quality due to compression, or limited by both. |
