Graph-based supervised discrete image hashing

摘要

Learning based hashing have been widely adopted to the approximate nearest neighbour search in large-scale image retrieval. However, how to preserve the semantic information in hashing embedding is still a challenge problem. Moreover, most of the existing methods employ the relaxation strategy to solve discrete constraint problem, which may accumulate binary quantization error as the coding length increases. In this paper, we propose a graph-based supervised hashing framework to address these problems, where the semantic information is preserved from two aspects. On one hand, we employ a supervised learning model to keep the semantic consistency. On the other hand, the intrinsic manifold structure is captured by a graph-based model. In addition, to reduce the quantization error, we adopt a discrete optimization strategy to replace the relaxation one. Experiments conducted on three benchmark datasets to demonstrate the effectiveness of the proposed method. (C) 2019 Elsevier Inc. All rights reserved.