Approximate Nearest Neighbor Search Based on Hierarchical Multi-Index Hashing

Nowadays, hashing methods are widely used in large-scale approximate nearest neighbor search due to its efficient storage and fast retrieval speed. By these methods, the original data is usually hashed into binary codes which enables to measure the similarity by Hamming distance. When dealing with large-scale data, their binary codes can be used as direct indices in a hash table. However, codes longer than 32 bits are obviously not efficient. For the given binary codes, this paper proposes a multi-index hashing structure based on binary code substrings partitioning. Since substrings partitioning is essentially a combinatorial optimization problem, we propose a hierarchical and recursive partitioning approach to obtain an approximate solution to it. Furthermore, we adopt a query-adaptive fine-grained ranking approach in the neighbor search stage to alleviate the imbalance between multi-index tables. Finally, extensive experiments are conducted on two datasets MNIST and CIFAR-10, demonstrating that our method achieves state-of-the-art performance in terms of efficiency, precision and recall rate.