Efficient Supervised Graph Embedding Hashing for large-scale cross-media retrieval

Abstract

Recently, graph based hashing has gained much attention due to its effectiveness in multi-media retrieval. Although several graph embedding based works have been designed and achieved promising performance, there are still some issues that need to be further studied, including, (1) one significant drawback of graph embedding is its expensive memory and computation cost caused by the graph Laplacian matrix; (2) most pioneer works fail to fully explore the available class labels in training procedure, which generally makes them suffer from unsatisfactory retrieval performance. To overcome these drawbacks, we propose a simple yet effective supervised cross-media hashing scheme, termed Efficient Supervised Graph Embedding Hashing (ESGEH), which can simultaneously learn hash functions and discrete binary codes efficiently. Specifically, ESGEH leverages both class label based semantic embedding and graph embedding to generate a sharing semantic subspace, and class labels are also incorporated to minimize the quantization error for better approximating the generated binary codes. In order to reduce the computational sources, a well-designed intermediate terms decomposition is proposed to avoid explicitly computing the graph Laplacian matrix. Finally, an iterative discrete optimal algorithm is derived to solve above problem, and each subproblem can yield a closed-form solution. Extensive experimental results on four public datasets demonstrate the superiority of the proposed approach over several existing cross-media hashing methods in terms of both accuracy and efficiency.