丛流形学习及其在物体识别中的应用

项目名称： 丛流形学习及其在物体识别中的应用

项目编号： No.61005004

项目类型： 青年科学基金项目

立项/批准年度： 2011

项目学科： 金属学与金属工艺

项目作者： 李春光

作者单位： 北京邮电大学

项目金额： 7万元

中文摘要： 流形学习作为高维数据分析方面一个活跃的研究方向，在机器学习、模式识别、机器视觉、信息检索和数据挖掘等领域具有重要的理论意义和潜在应用价值。近年来，随着对其本质的理解深入，其多方面缺陷被发现：单一流形的假设表达能力不足，需要数据良好采样，缺乏典型的成功应用等。因此，本项目拟从理论和应用两方面开展研究工作：(1)在理论方面，拟修改流形学习基本假设，使之能刻画存在复杂内在几何结构的高维数据集，并基于新的假设研究高维数据集上精细结构的无监督发掘问题；(2)在应用方面，针对具有多流形共存等复杂内蕴几何结构的高维数据集，研究无监督和半监督学习算法及其在视觉物体识别等方面的应用等。概括起来，本课题拟在现有流形学习理论基础上，修改数据集内蕴几何的基本假设，研究面向多流形共存等存在复杂几何结构的数据集的丛流形学习框架及其在图像物体识别中的应用，探索面向存在复杂内蕴结构的高维数据的新的分析工具。

中文关键词： 丛流形学习;物体识别;半监督学习;维数约减;高维数据分析

英文摘要： Manifold learning as an active research direction in the domain of high dimensional data analysis, has important theoretical significance and potential applications in machine learning, pattern recognition, machine vision, information retrieval, data mining and other fields. In recent years, with the deep understanding of its nature, many defects are found: (a) The single manifold assumption is lack of representation power, (b) It depends on well-sampled data (i.e. sufficient and uniform sampling), and (c) There is lack of typical successful applications. Therefore, in this project we intend to carry out our research from both theoretical and applied aspects: (a) In theoretical level, to change the old basic assumption in manifold learning, so that it can describe the intrinsic geometry of complex high-dimensional data sets, and to investigate the unsupervised learning framework to discover subtle structure based on the new assumptions; (b) In application level, to investigate unsupervised or semi-supervised learning algorithms for high-dimensional data with complex intrinsic geometric structure and its applications in Visual Object Recognition (VOC) and etc. To sum up, in this project we intend to extend the basic assumption beyond a single manifold to capture the complex intrinsic geometry in data, to investigate the bundle manifold learning framework and its applications in VOC, and to explore novel data analysis tools for high-dimensional data with complex intrinsic structure.

英文关键词： Bundle Manifold Learning; Visual Object Categorization; Semi-Supervised Learning; Dimensionality Reduction; High Dimensional Data Analysis