The objective of this project is to design efficient indexing strategies that support flexible retrieval metric through relevance feedback learning. However, trying to satisfy both goals (efficiency and flexibility) at the same time leads to a conflict. A novel approach is explored to capture the inherent interplay between flexible metrics and indexing that has the potential to resolve the conflict. It is hypothesized that the interplay can be exploited to create effective content-based retrieval systems that meet performance and computational challenges encountered in practical image database applications. This exploratory project seeks to establish the proof of concept of an approach that trades off accuracy for efficiency, and that can avoid exhaustive search in large-scale image databases. The methods to be explored are based on bump-hunting in high-dimensional data for inducing a set of (possibly overlapping) boxes that capture the local data distributions. The induced boxes effectively cover the feature space, thereby providing an index to the image database. The flexibility and efficiency of the novel indexing technique will be tested in heterogeneous image databases that support a variety of query types, ranging from query-by-image to query-by-region. If successful, the results of this project will enable the use of flexible metric learning in large scale image databases, which will have a significant impact in content-based image retrieval in broad areas such as health-care, scientific images, education, or art.
|Effective start/end date||11/1/01 → 11/30/02|
- National Science Foundation: $50,000.00