A Multidimensional Index Structure for Fast Similarity Retrieval
一种支持快速相似检索的多维索引结构

A typical example of similarity search is to find the images similar to a given image in a large collection of images. This paper focuses on the important and technically difficult case where each data element is represented by a point in a large metric space. As distance function employed is metric and distance calculations are assumed to be computationally expensive, it is necessary to index data objects in the metric space such that less distance evaluations are performed to support fast similarity queries. Based on the optimal partition method that uses representative points to partition the data space into subsets in a hierarchical manner, a novel distance-based index structure opt-tree and its variant h-tree are proposed. In order to fully support the content-based image retrieval, the optimal strategies for the partition of data space and data redundancy storage, which are called h-optimal partitioning and h-symmetric redundancy storage respectively, are adopted in the h-tree index structure to achieve the high performance of the similarity retrievals. In this paper, the decisions and the algorithms which led to opt-tree and its variant h-tree are discussed in detail, and the experimental results show that this index structure is effective.