Two-Layer Localized Sensitive Hashing with Adaptive Re-Ranking

Hashing-based approximated nearest neighbor (ANN) search techniques have been widely studied owing to its compact binary codes and efficient search scheme for large-scale image retrieval. For the most popular existing hashing methods, e.g. the Locality sensitivity Hashing and the Spectral Hashing, the key issue is to choose appropriate binary code length for similarity preserving and computational efficiency. Several extensions have been proposed to address the problem of balancing precision, recall rate and computation efficiency. However, most of existing hashing methods struggle for how to choose appropriate length of hash codes. A bad choice of code length may result in extremely poor performance of retrieval. In this paper, we propose a novel hashing scheme, called the Two-Layer Localized Sensitive Hashing with Adaptive Reranking (TL-LSHAR). This method utilizes a short hash code to generate the weights for a long hash code to further improve the retrieval performance. Moreover, the new scheme can be used by most of the existing hashing methods. The performance is evaluated on two large scale image databases which demonstrates the efficiency of our scheme.