A semantic object-oriented model for content-based retrieval.

摘要

Multimedia data is becoming the primary data type for many application domains. Furthermore, now multimedia devices such as scanners, digital cameras, and microphones are becoming cheap and readily available, meaning that the use of multimedia data will continue to grow. However, the growth of multimedia data has not been paralleled by growth in the ability to manipulate and manage the data.; Current data management techniques have been based primarily on techniques learned from managing alphanumeric data. However, multimedia data is significantly different than alphanumeric data in two fundamental ways. First, a single multimedia object is typically large meaning that traditional alphanumeric database storage techniques are inappropriate. Second, multimedia data is generally meaningless to a human. Looking at multimedia data does not give any clues as to what that data means and/or contains.; Research is currently underway to develop new techniques for storing, searching, and retrieving multimedia data. Most multimedia content-based retrieval tools provide capabilities to search for features. However, multimedia data, unlike alphanumeric data, is used primarily to convey conceptual information that is evident when the data is taken as a whole.; Developing a multimedia database system that automatically identifies conceptual content is difficult. Content identification is complicated by two primary factors. First, as stated earlier, multimedia data tends to be large, meaning that it can take considerable processing to identify a single feature. Second, given our current understanding of content identification, it is impractical to build a single system that can identify content in any domain.; This thesis describes a new model for managing multimedia data called MOODS. MOODS takes a novel approach to the design of information management systems that incorporate the ability to extract both the basic and high-level semantic concepts, with the ability to directly model and manipulate multimedia data objects. We identify four major contributions. First, MOODS provides the ability to create domain-specific multimedia databases. Tailoring a database to a specific domain allows the content-retrieval engine to focus exclusively on the content important in that domain.; Second, a MOODS multimedia database provides automatic access to the full range of semantic content in multimedia. A MOODS database incorporates processing routines for identifying basic “low-level” features. Furthermore, a MOODS database also includes a knowledge base for modeling and representing the high-level semantic concepts that cannot be identified through processing alone.; Third, because identifying content can take a considerable amount of time, a MOODS database uses a combination eager/lazy approach to content identification. Based on a user's perception of what content is important and what is not, a MOODS database will automatically identify all the “important” content in a user's data, and postpone identifying less-important content until it is actually needed.; Fourth, MOODS uses a novel semantic object-oriented language that has use beyond multimedia databases. In the MOODS language, objects contain data and functions like conventional objects, but they also include a semantic description that describes what is currently known about the content of the object. (Abstract shortened by UMI.)