基于字典学习的跨媒体检索技术

Cross-media retrieval technology based on dictionary learning

在研究跨媒体信息检索时,对于不同模态数据的异构性提出了挑战,针对如何更好地克服异构问题以提高多模态数据之间的检索精度,提出了一种基于字典学习的跨媒体检索新技术。首先,通过字典学习方法学习两个不同模态数据之间的稀疏系数;然后,通过特征映射方案由两个不同的投影矩阵分别把它们投入共同的特征子空间;最后,通过标签对齐同一类来增强不同模态之间的相关性。实验结果表明,与传统的同构子空间学习方法相比,基于字典的算法分类性能优越,该实验方法在两个数据集上优于几种最先进的方法。

In the study of cross-media retrieval,how to capture and correlate heterogeneous features originating from different modalities remains a challenge.To cope with the aforementioned problems,this paper presented a novel cross-modal retrieval framework based on coupled dictionary learning.Firstly,it obtained sparse coefficients from different modalities by imposing dictionary learning.Then,it projected the data samples from different modalities into a common feature space.Moreover,it leveraged label information to align the cross-modal data sample pairs in the common space so as to encourage the inherent correlation across the different modalities.Simulation experimental results show that the method based on dictionary learning algorithm has superior recognition performance in comparison with the methods based on traditional mid-level feature subspace.Experiment results on two public datasets demonstrate that this method outperforms several state-of-the-art methods.