Discrete Multi-graph Hashing for Large-Scale Visual Search

Hashing has become a promising technique to be applied to the large-scale visual retrieval tasks. Multi-view data has multiple views, providing more comprehensive information. The challenges of using hashing to handle multi-view data lie in two aspects: (1) How to integrate multiple views effectively? (2) How to reduce the distortion error in the quantization stage? In this paper, we propose a novel hashing method, called discrete multi-graph hashing (DMGH), to address the above challenges. DMGH uses a multi-graph learning technique to fuse multiple views, and adaptively learns the weights of each view. In addition, DMGH explicitly minimizes the distortion errors by carefully designing a quantization regularization term. An alternative algorithm is developed to solve the proposed optimization problem. The optimization algorithm is very efficient due to the low-rank property of the anchor graph. The experiments on three large-scale datasets demonstrate the proposed method outperforms the existing multi-view hashing methods.