Semantic video model for content-based retrieval

Traditional research on video data retrieval follows two general approaches. One is based on text annotation and the other on content-based comparison. However these approaches do not fully make use of the meaning implied in a video stream. To improve these approaches, a semantic video model cooperating with a knowledge database is studied. We propose a new semantic video model and focus on presenting the semantic meaning implied in a video. According to the granularity of the meaning implied in a video, a five-level layered structure to model a video stream is proposed. A mechanism is also provided to construct the five levels based on the knowledge categories defined in the knowledge database. The five-level layered structure consists of raw-data levels and semantic-data levels. A uniform semantics representation is proposed to represent the semantic-data levels. This uniform semantics representation allows measuring the similarity of two video streams with different duration. Then an interactive interface can provide browsing and querying video data efficiently through the uniform semantics representation.