智能感知与计算研究中心知识库

With the rapid growth of visual data on the web, deep hashing
has shown enormous potential in preserving semantic similarity for visual search. Currently, most of the existing hashing methods employ pairwise or triplet-wise constraint to obtain the semantic similarity or relatively similarity among binary codes. However, some potential semantic context cannot be fully exploited, resulting in a suboptimal visual search.
In this paper, we propose a novel deep hashing method, termed Joint Multiply Semantics Hashing (JMSH), to learn discriminative yet compact binary codes.
In our approach, We jointly learn multiply semantic information to perform feature learning and
hash coding.
To be specific, the semantic information includes the pairwise semantic similarity between binary codes, the pointwise binary codes' semantics and the pointwise visual feature' semantics.
Meanwhile, three different loss functions are designed to train the JMSH model. Extensive experiments show that the proposed JMSH yields state-of-the-art retrieval performance on representative image retrieval benchmarks.