Improving high-dimensional indexing for content-based image retrieval.
详细信息
文摘
Most high-dimensional indexing schemes proposed for similarity query in content-based image retrieval (CBIR) systems are tree-structured. The quality of a high-dimensional tree-structured index is mainly determined by the algorithm of effectively inserting feature vectors which represent multimedia objects into the index tree. In this dissertation, the issues related to high-dimensional indexing for content-based image retrieval are investigated.;A heuristic-based approach in the tree-descending phase during insertion is developed. A beam search algorithm is incorporated into the insertion algorithm. Furthermore, several node-splitting strategies for tree-structured high-dimensional indexes are studied. Moreover, by building different indexes with different features, the relative significance of different image features in content-based indexing and retrievals is explored. Finally, the quality of using different combinations of feature sets for CBIR is studied.;The developed indexing scheme with the insertion and search algorithms has been implemented and extensively tested with two image databases. The experimental results show that the methods can be applied to improve high-dimensional tree-structured indexes.