基于关系矩阵融合的多媒体数据聚类

Multimedia Data Clustering Based on Correlation Matrix Fusion

针对目前多媒体聚类研究中如何挖掘和利用不同数据集之间统计关系的问题,提出一种基于关系矩阵融合的聚类方法,首先,对图像和音频数据集中提取的特征矩阵进行相关性分析和子空间映射,进而在全局范围内对图像相似度、音频相似度以及图像和音频的相关度进行融合与优化,最后,采用基于相似度的循环迭代算法进行图像和音频聚类.对比实验从多个角度验证了文中方法的有效性,并能较好地应用于多媒体交叉检索.

It is a hot issue to explore statistical correlation between different types of multimedia data, especially in the area of multimedia clustering. In this paper, we propose a multimedia clustering method based on correlation matrix fusion. Visual and auditory feature matrices are firstly initialized and simultaneously mapped into a subspace; Then we utilize correlation fusion strategy on image similarity matrix, audio similarity matrix and image-audio correlation matrix for global reinforcement and optimization; Thirdly, similarity-based clustering method is implemented for image and audio clustering in the subspace. Experiment results are encouraging and show that the performance of our approach is effective. Besides, an interesting experiment of image-audio crossretrieval validates the applicability of our approach.