|Title||Advances in Sprite-based Video Coding - Towards Universal Usability|
|Tutor||Prof. Dr. Thomas Sikora, Prof. Dr. Jörn Ostermann (Leibniz Universität Hannover)|
|Abstract||This dissertation presents new approaches end extended techniques for the coding of digital video using background sprites, also called background mosaics. Sprites form a visual summarization of the rigid background of a captured scene shot. They are represented in oversized images, which preferably do not contain any foreground objects. This type of redundancy reduction is an ideal tool for video coding since the complete background information can be stored in the sprite image and some additional projection parameters. However, the generation of sprites is only possible for certain scenes. Since a successful coding strategy has to be universally applicable, the development of techniques for facilitating a broader use of sprite-based video coding represents the main focus of this thesis. Early approaches, as the one adopted in the MPEG-4 standard, have not been utilized due to the lack of universality and usability.|
For this purpose, we present techniques for the generation of multiple sprites and provide automatic segmentation approaches for the independently moving foreground objects. While multiple sprites prevent the construction of degenerated sprites and simultaneously minimize the impact of geometrical distortions, the segmentation enables the automatic discrimination in foreground and background objects. Thus, it is a fundamental tool for object-based video coding. The presented segmentation techniques are built upon the background sprites and thus, are easy to integrate into the overall coding process.
The improvement of the background modeling using sprites marks another important aspect of this dissertation. Since state-of-the-art hybrid coding strategies work very efficient and yield high quality results, the prediction quality of the background using sprites has to be improved remarkably. In order to achieve this goal, we present novel image registration and sprite generation algorithms. Especially the potential of super-resolution processing will be exploited. Due to the capturing process, we obtain several differently sampled versions of the same image content. This fact can be used for the construction of background sprites of enhanced resolution, which has a positive influence on the resulting coding quality as well as on the rate-distortion results.
Eventually, two techniques for sprite-based video coding are presented. Both approaches utilize above mentioned tools for improving quality and universality of the sprites. The coding gain over latest standards proves their usefulness. A complete coding system for the processing of any video content is still not achieved, but an outlook of its possible architecture is drafted. Thus, this thesis contributes to a gradual change of the video coding paradigm, where additional instruments from computer vision and computer graphics are utilized to unequally encode independent parts of a video scene.
|Key words||Sprite generation, Mosicing, Super-resolution, Video Coding, Object Segmentation, Multiple Sprites|