A Discriminative Framework for Hashing

Abstract

Hashing methods have been widely applied for fast retrieval and efficient data storage. However, most existing hashing methods have not taken the discriminative features into account in nearest neighbors search which leads to unsatisfied retrieval accuracy, especially for high-dimensional dataset. In order to take best use of discriminative information, we introduce an effective feature extraction framework for hashing that can get high retrieval accuracy in this paper. Firstly, the divergence between two classes is represented by the ratio of signal to noise. Then the discriminative features of high-dimensional data are obtained through the generalized eigen-decomposition. Finally, we exploit the discriminative feature into hashing methods to generate compact binary code. Experimental results on data sets show that the proposed framework can reach better results in comparison with state-of-the-art methods.