Hierarchical soft clustering tree for fast matching of binary codes

Abstract

Recent studies show that binary codes are a powerful means to perform large-scale image searching and feature matching. Binary codes require significantly less storage space while enabling efficient computation in the Hamming space. At the core of the algorithms that utilize binary codes is the nearest neighbor problem. Although the distance in the Hamming space can be computed efficiently, a linear search against a large-scale dataset creates a bottleneck. Some approximate nearest neighbor (ANN) search algorithms have been proposed to overcome this limitation

the hierarchical clustering tree (HCT) was proved to outperform the well-known locality sensitive hashing. HCT recursively clusters the input binary codes by assigning each binary code to only one cluster, which leads to a quantization error. As a solution to this problem, in this thesis, we propose two algorithms to create hierarchical soft clustering tree (HSCT) by assigning a data point to multiple nearby clusters in the Hamming space. One is a fuzzy hierarchical soft clustering tree (FHSCT) which soft assigns a data point to multiple nearby clusters based on fuzzy membership weight without considering an expected query point. To increase search precision, the other algorithm, query-aware hierarchical soft clustering tree (QAHSCT) is proposed. QAHSCT assigns a binary code to a cluster based on the probability that the query point nearest to the binary code is assigned to the cluster. Because query points are not available when building trees, the exact distance between a query point and cluster center cannot be obtained. Therefore, distance distribution is used instead, which is approximated using Poisson binomial distribution. Experiments show that our proposed methods outperform the previous ANN search methods in terms of search speed without sacrificing search precision.