Negotiating the semantic gap in an MPEG-7 aerial image database

This dissertation presents our research related to content-based retrieval and annotation system for aerial image regions in a database environment. The purpose is to narrow the semantic gap for content-based image retrieval (CBIR) applications in a dynamic interaction context and provide a flexible content description framework in terms of MPEG-7 descriptors through the re-clustering process. Our work contributes to the following aspects: first, we utilize a split and merge process for image segmentation, which is more relevant to human perception, and integrate quadratic distance with latent semantic indexing technique to negotiate semantic gap; second, we illustrate a process to extract emergent semantics in a dynamic context, for example, the system can learn patterns from the users' interaction to perform re-classification and re-segmentation; third, we integrate different MPEG-7 descriptors in a unified framework for our CBIR application in database environment, which not only provides convenience for user interaction, but also improves the performance of our application.; In order to show its effectiveness, we designed a content-based aerial image region retrieval and annotation system, and promising results were obtained. A series of innovative components was developed in our system, of which the application is not restricted to this particular system, but can be used to solve many more general and complicated problems. For example, we built a flexible framework with MPEG-7 descriptors, which was used to support the users in extracting emergent semantics in a dynamic context. This new approach can be widely used in a re-clustering process for CBIR applications. Besides, we explored building an index to support XQuery and XUpdate for a CBIR system in a native XML database environment. Our experience showed we can utilize classification information under certain similarity metric to build an efficient index for content-based retrieval and update.; Beyond our system, we also propose to integrate different MPEG-7 descriptors in different stages of CBIR applications, building an efficient platform to narrow semantic gap, as we often can only extract semantics in a dynamic query and annotation context.