基于映射字典学习的跨模态哈希检索

Projective Dictionary Learning Hashing for Cross-modal Retrieval

针对网络上出现越来越多的多模态数据,如何在海量数据中检索不同模态的数据成为一个新的挑战.哈希方法把数据映射到Hamming空间,大大降低了计算复杂度,为海量数据的跨模态检索提供了一条有效的路径.然而,大部分现存方法生成的哈希码不包含任何语义信息,从而导致算法性能的下降.为了解决这个问题,本文提出一种基于映射字典学习的跨模态哈希检索算法.首先,利用映射字典学习一个共享语义子空间,在子空间保持数据模态间的相似性.然后,提出一种高效的迭代优化算法得到哈希函数,但是可以证明问题的解并不是唯一的.因此,本文提出通过学习一个正交旋转矩阵最小化量化误差,得到性能更好的哈希函数.最后,在两个公开数据集上的实验结果说明了该算法优于其他现存方法.

With the sharp increasing of multi-modal data on the internet, retrieving samples with different modalities has become a challenge, hashing methods, which map the data to Hamming space to reduce computational cost, can provide an effective way for large-scale cross-modal retrieval. However, the hashing codes of most existing methods do not contain any semantic information, which degrades the performance. To address this issue, we propose a cross-modM hashing method, termed projective dictionary learning hashing （PDLH） method. Firstly, projective dictionary learning is employed to learn a sharing semantic subspace by preserving the inter-modal similarity. Then an efficient iterative optimal algorithm is proposed to gain hashing hmctions. However, the solution is not a unique solution, as proven in this paper. To further improve the performance of the proposed method, an orthogonal rotation matrix is learned by minimizing the quantization loss for better hashing functions. Finally, experimental results on two widely used datasets show that the performance of the proposed method is better than those of the existing methods.