高维多媒体特征的低维流形子空间降维及聚类研究

项目名称： 高维多媒体特征的低维流形子空间降维及聚类研究

项目编号： No.61472172

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 岳峻

作者单位： 鲁东大学

项目金额： 84万元

中文摘要： 跨媒体与单一媒体相比更加符合人类大脑对信息的认知处理模式，然而不同类型半结构化和无结构化媒体数据特征的高维度和差异性对其使用造成了困扰。多媒体特征降维与聚类分析的目标是克服这一问题，实现跨媒体数据的有效利用。本课题以网络数据中人的文本、图像、视频、音频和三维模型这5种类型媒体形态数据为研究对象，研究建立异构媒体相关性特征选择与映射方法，探明不同类型媒体底层到高层特征的有效映射机制；研究基于流形学习的异构媒体语义特征空间降维方法，建立高维媒体特征的低维语义子空间，解决图像、视频等媒体特征维数过高引起的维数灾难问题；研究基于核函数模糊聚类的子空间特征相似度计算模型，解决不同模态媒体特征在低维非线性特征子空间中的相似度计算问题；最后对所提方法进行验证、评估与修正。通过本课题研究为实现跨媒体数据高效综合利用提供理论与实验依据。

中文关键词： 跨媒体；流形学习；模糊聚类；语义特征

英文摘要： Cross-media is more in line with the information cognitive of human brain than single media. However, the high-dimensional and differences of different media features caused the disorder application of cross-media data. The features dimensionality reduction and clustering is proposed to make unified understanding of different semi-structured or non-structured media in cyberspace. In this study, we will research the human beings information in cyberspace with five different types of media mode including text, image, video, audio and 3D model. The features correlation selecting and mapping model will be built in order to explore the efficient mapping mechanism between low-level and high-level features. The manifold learning based semantic features space dimensionality reduction will be studied to solve the curse of dimensionality problem of image, video and other media features. The kernel-based fuzzy clustering method will be researched to build the similarity computing model of different media features in the non-linear manifold subspace, which can solve the problem of similarity computing of different media data in low-dimensional feature spaces. After these studies, the proposed methods will be verified and modified. The researches can provide a theoretical and practical basis for the high-efficiency application of cross-media data.

英文关键词： cross media;manifold learning;fuzzy clustering;semantic feature