VP-EMD Tree: An Efficient Indexing Strategy for Image Retrieval

In order to utilize the large volume of images in databases efficiently, content-based image retrieval (CBIR) has been successfully employed whereby instead of the actual images extracted feature descriptors are used. CBIR systems usually employ multidimensional indexing structures, such as the R*tree or the SR-tree to index these features and thus to speed-up the query performance. However, the majority of these indexing structures are only suitable for features with fixed length since they mainly employ partitioning methods to divide the multidimensional vector space into sub-spaces. These structures are not appropriate for features with varying characteristic, such as unspecified order of the vector elements and variable lengths. Data with weight information cannot be indexed via current structures either. Hence, this paper introduces a novel indexing structure, the VPEMD tree. It overcomes the limitations of traditional multidimensional indexing structures and speeds-up the query performance for the features under consideration significantly.