Speeding up the similarity search by contractive embedding methods

本文的研究目的是开发一种高效的多媒体数据（如图片图像和电影）相似度搜索方法。虽然Bustos等人设计了几种基于枢轴的方法，并取得了优异的结果，但在这一领域仍有进一步研究的空间。我们设计了两种基于轴心的相似度搜索方法。第一种方法的特点是改进了busstos的主元交换算法。第二种方法是基于曼哈顿距离的广义枢轴法。我们通过扩展我们以前开发的一种方法来设计它。实验结果表明，这两种方法在检索性能和创建时间上都优于传统方法。为了更有效地阐明多媒体数据的本质，我们设计了网络社区的提取方法和网络可视化方法。

The purpose of our research is to develop efficient similarity search methods of multimedia data such as picture images and movies. Although Bustos and others have devised several pivot-based methods and have obtained the excellent results, in this field there is still room for further investigation.We devised two pivot-based similarity search methods. The first method is characterized by improving Bustos's exchange algorithm of pivot set elements. The second method is a generalized pivot method based on Manhattan Distance. We devised it by extending one of our previously-developed methods. By experiments, we showed that both methods were better than the conventional methods in retrieval performance and creation time of pivot sets.Furthermore, we devised extraction methods of the communities in networks and network visualization methods in order to make clarifying the nature of the multimedia data more efficient.