Locality preserving hashing

Abstract

The spectral hashing algorithm relaxes and solves an objective function for generating hash codes such that data similarity is preserved in the Hamming space. However, the assumption of uniform global data distribution limits its applicability. In the paper, we introduce locality preserving projection to determine the data distribution adaptively, and a spectral method is adopted to estimate the eigenfunctions of the underlying graph Laplacian. Furthermore, pairwise label similarity can be further incorporated in the weight matrix to bridge the semantic gap between data and hash codes. Experiments on three benchmark datasets show the proposed algorithm performs favorably against state-of-the-art hashing methods.