标签局部结构保持的离散哈希方法

Discrete Hashing Based on Label Local Structure Preserving

跨模态哈希检索以其存储消耗低、检索效率高引起广泛关注.现有研究仍存在如何保持标签语义和语义信息损失问题.为此,本文提出了一种新颖的标签结构保持离散哈希方法LSPDH(Label Structure Preserving Discrete Hashing),该方法将哈希学习分为哈希码学习和哈希函数学习两步.哈希码学习中,不同于其他使用流形学习提取多模态数据特征的方法,本文方法则运用流形学习提取标签信息的局部结构并在汉明空间中保持,同时将标签信息映射到哈希码矩阵,融入哈希码学习过程,来降低构建流形相似度矩阵所造成的语义信息损失;哈希函数学习中使用了核函数来获取数据间的非线性关系.最后,通过三个基准数据集上与近期跨模态哈希方法对比实验,验证了提出方法的有效性.

Cross-modal hashing attracted many people′s attention in cross-modal retrieval due to its characteristics of low storage consumption and high retrieval efficiency.However,many existing methods still have problems of how to preserve label semantic and reduce semantic information loss.To address these issues,a novel method LSPDH(Label Structure Preserving Discrete Hashing)is proposed in this paper,LSPDH is divided into two steps:hash codes learning and hash functions learning.In hash codes learning,different from other methods that use manifold learning to extract multimodal data features,LSPDH leverages the manifold learning to preserve the local structure of label information into the Hamming space,meanwhile the label information is mapped to the hash codes and integrated into the hash codes learning process to reduce the semantic information loss caused by the construction of the manifold similarity matrix.In hash function learning,kernel function is used to catch the nonlinear relationship between data.The experimental results on LSPDH and several state-of-the-art cross-modal hashing methods on three benchmark datasets demonstrate the effectiveness of LSPDH.