Video Processing Using Spatiotemporal Structure

141 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2003.This thesis focuses on how to identify the interframe redundancy in a video sequence by using hierarchical multiscale segmentation and multiresolution motion representation. The multiscale hierarchy in a video sequence is treated as image properties over time. The spatial structure of individual frames is extracted as the multiscale layout of regions. Such spatial structure is related across the frames to estimate temporal structure or motions of regions. The structural estimates are then utilized for both video processing using local spatiotemporal structure models of motion boundaries and video coding using backward multiscale (i.e., multiresolution) motion representations. Models of local spatiotemporal structure are exploited for efficient block-based motion estimation and frame interpolation. With these models, boundaries of moving objects (foreground and background) are identified using predictive region segmentation and a mesh topology. The multiscale image segmentation is used to develop a backward region based video compression algorithm for low bit-rate applications. The selection of an optimal scale of segmentation from the view point of compression is proposed, and the occlusion problem is tackled as well. The performance of residual coding is improved by using the fact that energy of the residual resulting from motion compensation is concentrated in a priori predictable positions. A spatially scalable backward wavelet video coding algorithm is presented, in which a computationally efficient algorithm to generate overcomplete wavelet coefficients is proposed. Wavelet coefficients of high bands are predicted without an interpolation operation from the coarsest level in a hierarchical backward fashion.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD