Abstract
In recent years, the revolution in computational photography has appeared with the invention of light field cameras. The specific ability of changing the focal planes after the light field is captured enables users to edit their shallow-depth-of-field pictures at any time. Despite the fact that many research works achieved advanced improvement on light-field processing, high-resolution and realistic refocusing is still a challenge. Therefore, it is an essential issue to resolve for enabling high-quality light-field video applications. In this thesis, we propose a block-based processing structure for refocusing for sparse light fields. Using sparse light fields effectively reduces the storage requirement and achieves higher image resolution compared to dense light fields. We devise a timing-based block merging algorithm to achieve the best quadtree partition in terms of computation time. Also, we implement a multi-level pyramid refocusing scheme to reduce the unnecessary computation, especially for the blocks with heavily defocused blurs. However, overlapped blocks are required to eliminate the block effect, and we proposed a fast block-border search method to select the boundaries efficiently. Finally, we apply our algorithm to many light fields which are from different sources and have different characteristics. In our experiment, we show that the speedup performance is scene-dependent and can be up to 21x without visible quality degradation.