In image databases, efficiency of processing similarity queries for content-based image retrieval is very important. For this reason, many high-dimensional index structures for similarity queries have been proposed in literature. Among these, index structures based on a data-partitioning method, e.g., R*-tree, X-tree, TV-tree, and SR-tree, are widely used. In this approach, the entire data space is managed by examining distribution of data objects being inserted into the tree. Since similar data objects are clustered into the same data cluster, data-partitioning approach is suitable for processing similarity queries.
However, data-partitioning index structures suffer from performance degradation as the dimensionality of data objects increases. This problem is caused by improper management of data clusters at insertion time. So, we propose a new clustering method based on two measures: DOC(Degree of Coherence) and DOA(Degree of Affinity). Here, DOC is a degree of similarity among data objects within a data cluster and, DOA is a degree of similarity among data objects belonging to differnt data clusters. Our experiments show that the proposed clustering method improves the efficiency of similarity queries processing.
