Learning robust graph hashing for efficient similarity search

Unsupervised hashing has recently drawn much attention in efficient similarity search for its desirable advantages of low storage cost, fast search speed, semantic label independence. Among the existing solutions, graph hashing makes a significant contribution as it could effectively preserve the neighbourhood data similarities into binary codes via spectral analysis. However, existing graph hashing methods separate graph construction and hashing learning into two independent processes. This two-step design may lead to sub-optimal results. Furthermore, features of data samples may unfortunately contain noises that will make the built graph less reliable. In this paper, we propose a Robust Graph Hashing (RGH) to address these problems. RGH automatically learns robust graph based on self-representation of samples to alleviate the noises. Moreover, it seamlessly integrates graph construction and hashing learning into a unified learning framework. The learning process ensures the optimal graph to be constructed for subsequent hashing learning, and simultaneously the hashing codes can well preserve similarities of data samples. An effective optimization method is devised to iteratively solve the formulated problem. Experimental results on publicly available image datasets validate the superior performance of RGH compared with several state-of-the-art hashing methods.