Statistical semantic analysis of spatio-temporal image sequences
Ying Luo, Jenq–Neng Hwang · 2004
Semantic analysis to obtain the content description of spatio-temporal image sequences such as video clips and medical image sequences is the critical task to facilitate the indexing, retrieving and filtering of the image sequence databases. This thesis proposes a novel statistical framework based on Bayesian statistics to model the image sequences. The algorithms of different dynamic Bayesian networks as applied to various spatio-temporal image sequences are discussed. The advantages and drawbacks are presented. A novel fine-level object-based analysis scheme is proposed for different image sequences. A key factor about the image sequences is that their contents are defined on various levels, from coarse to fine. An innovative coarse-to-fine framework, which provides an excellent framework that guides the hierarchical analysis of the complicated image sequences, is presented. It is proved that the coarse-to-fine analysis is a Bayesian methodology and the dynamic Bayesian networks and other methodologies can be naturally glued together through the coarse-to-fine analysis to yield a complete description of the image sequence contents. The modeling of magnetic resonance image sequences and video image sequences are systematically studied under the proposed framework. Both the coarse-level and fine-level analysis are performed separately and finally put together. It is shown that framework is effective and efficient in obtaining the complete content description of complicated spatio-temporal image sequences.