Weakly supervised locality sensitive hashing for duplicate image retrieval

Locality sensitive hashing (LSH) is quite popular in high dimensional data indexing. However, most of existing methods perform hashing in an unsupervised way, that is to say, hash functions are randomly generated without the prior information of the data. In this paper, we propose two improved LSH algorithms based on weakly supervised learning technique, which need only small quantities of labeled sample pairs. One is to select the most appropriate hash functions from a pool of functions using sample pairs labeled with “similar” or “dissimilar”. The other is to generate hash functions with positive sample pairs. The experiments show that the proposed algorithms reduce the search complexity compared with original LSH.