基于流形学习理论的图像搜索技术研究

项目名称： 基于流形学习理论的图像搜索技术研究

项目编号： No.60875044

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 机械、仪表工业

项目作者： 何晓飞

作者单位： 浙江大学

项目金额： 35万元

中文摘要： 随着数码照相设备以及互联网的迅速发展，数字图像的数量呈几何式的增长。如何快速有效地管理、浏览、搜索图像信息成为一个具有挑战性的问题。传统的图像搜索算法利用用户的相关反馈去学习一个距离度量或者线性分类器，并对数据库中的图像进行排序。这些算法把数据空间看作是一个欧氏空间，而事实上图像信息所张成的空间通常是嵌在高维欧氏空间中的一个低维子流形。有效地学习这个图像子流形的几何拓扑结构，比如度量张量、本征维度、同调等等，将大大的提高图像搜索的性能。本项目将致力于流形学习的基础理论研究以及在图像搜索上的应用。通过提取图像的视觉特征，以及用户提供的相关反馈，图像子流形可以表示为一个连通图。利用谱图理论，流形的连续几何拓扑结构可以离散地在连通图上估计得到。通过正则化分析，学习到的流形结构将用于设计图像排序、分类和聚类算法，并在此基础上尝试构建新型的图像搜索引擎模型。

中文关键词： 图像搜索；流形；谱图理论；模式分类；正则化

英文摘要： Due to the rapid development of digital photograph equipments and the Web, there is exponential growth of the number of digital images. A challenging problem is how to efficiently and effectively manage, browse, and search the images. Traditional image retrieval algorithms make use of the users' relevance feedbacks to learn a distance metric or a linear classifier, and rank the images in the database. These algorithms consider the data space as a Euclidean space, but the images are possibly sampled from a lower dimensional submanifold which is embedded in high dimensional ambient space. The image retrieval performance can be significantly improved if we can learn the geometrical and topological structure of the submanifold, such as metric tensor, dimensionality, and homology. This project is devoted to theoretical development of manifold learning theory, as well as its application to image retrieval. By extracting visual features of the images and the user provided feedbacks, the image manifold can be modeled by a connected graph. Using spectral graph theory, the geometrical and topological structure may be discretely approximated from the graph. The learned manifold structure will be used to design novel image ranking, classification and clustering algorithms through regularization techniques, based on which we will construct a novel image retrieval system.

英文关键词： image retrieval; manifold; spectral graph theory; pattern classification; regularization