Fuzzy Clustering Method for Content-based Indexing

E cient and accurate information retrieval is one of the main issues in multimedia databases. In content-based multimedia retrieval databases, contents or features of the database objects are used for retrieval. To retrieve similar database objects, we often perform a nearest-neighbor search. A nearest-neighbor search is used to retrieve similar database objects with features nearest to the query under the feature vector space with a given distance function (similarity measurement). Typically, data exist in natural cluster. However, many of the currently indexing methods do not utilize this data cluster information in the construction of the indexing structure which leads to performance degradation. To improve the retrieval performance, we (1) use Fuzzy Competitive Clustering (FCC), a noise resistance fuzzy clustering algorithm, to locate good approximate cluster prototypes e ciently, (2) use the result of FCC clustering to construct a good indexing structure (FCC-btree) for e ective nearest-neighbor search and (3) Dervied two elimination rules for purning the indexing tree in searching process. Our experimental results show that: (1) FCC gets the better cluster prototypes then other traditional clustering algorithms in general. and (2) The FCC-b-tree always has a better performance than linear search.