多重稀疏特性的核子空间分析理论与应用

项目名称： 多重稀疏特性的核子空间分析理论与应用

项目编号： No.61271412

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 无线电电子学、电信技术

项目作者： 潘静

作者单位： 天津职业技术师范大学

项目金额： 70万元

中文摘要： 由于能通过核函数途径挖掘高维数据的内在非线性结构，核子空间分析已成为最有效的维数约简方法之一。但现有稀疏核子空间分析仅有单个稀疏特性，其作用仅限于减少核函数计算次数以提高特征提取速度。显然，现有核子空间分析缺乏多个重要的稀疏特性，甚至都不具备稀疏线性子空间分析已有的稀疏特性。稀疏特性的缺失极大限制了核子空间分析的局部性(抗遮挡)、抗过拟合性、鲁棒性、高效性和可解释性。为克服上述严重问题，本项目提出具备多重优良稀疏特性的核子空间分析理论，通过稀疏度转化和显式特征映射，使特征空间的甚高维投影向量本身，以及甚高维投影向量对应的投影系数也具有稀疏特性。还提出稀疏特征加权和维数约简统一在稀疏核子空间分析框架下同时进行的方法，比单独进行的方法能使总体目标达到更优并具有更强的鉴别力。由于赋予了多重优良的稀疏特性并显著提高特征提取和分类性能，本项目对核子空间分析理论发展及在视觉计算中的应用具有重要作用。

中文关键词： 特征提取；子空间分析；维数约简；核子空间分析；

英文摘要： Kernel subspace analysis is one of the most effective feature extraction methods. Using the kernel trick, it is powerful to discover the underlying intrinsic nonlinear structure of the high-dimensional data. However, existing kernel subspace analysis method possesses merely one sparsity property which is useful to fast feature extraction by reducing the number of computation of kernel functions. Obviously, existing kernel subspace analysis lacks of several important sparsity properties which even occur in sparse linear subspace analysis. This drawback leads the kernel subspace analysis not to be efficient, interpretable, and robust to occlusion， noise, and overfitting. To overcome the above disadvantages, we propose a unified kernel subspace analysis framework so that it has multiple desirable sparsity properties. By sparsity tranfer and explicit feature map, we make the very high dimensional projection vectors and the expansion coefficients in feature space are sparse. This project also proposes a sparse feature weighting method in the framework of kernel subspace analysis with the benefit of better discriminative ability. The proposed methods are expected to remarkably improve the feature extraction and pattern classification performances for high-dimensional data such as image and video. Therefore, the proje

英文关键词： feature extraction；subspace analysis；dimensionality reduction；kerenl subspace analysis；