面向特征提取的低秩与稀疏图嵌入理论与算法研究

项目名称： 面向特征提取的低秩与稀疏图嵌入理论与算法研究

项目编号： No.61503195

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 自动化技术、计算机技术

项目作者： 黄璞

作者单位： 南京邮电大学

项目金额： 22万元

中文摘要： 特征提取是处理高维数据、获得有效信息的重要途径。目前，基于图嵌入的特征提取方法是主流的特征提取方法，这类算法中，图的构建对于算法性能的好坏起着决定作用。低秩表示是一种重要的多子空间数据分析方法，其能够获得数据的全局结构，对于噪声具有极强的抗干扰性，因此在许多领域中得到了广泛应用。本项目针对特征提取中的构图问题，结合低秩与稀疏分解技术，提出以下几方面研究内容：(1)将低秩表示用于构造低秩、稀疏的图，并结合经典的特征提取方法设计新的特征提取算法；(2)研究图学习与特征提取同时进行的学习框架，以便能够利用不同投影空间中的数据信息；(3)以数据自身的特性及一些先验信息为指导，建立适用于特征提取的低秩表示模型。项目研究能够更有效地处理、分析高维数据，与传统的特征提取算法相比，在对噪声的鲁棒性、参数选择及特征提取的有效性等方面均有较大的改善，同时也为计算机视觉、模式识别领域的相关研究提供了理论支撑。

中文关键词： 子空间方法；鉴别子空间；投影；线性判别分析；鉴别向量

英文摘要： Feature extraction is an important way to deal with high-dimensional data and obtain effective information. So far, graph embedding based feature extraction methods have been widely employed for feature extraction. In these methods, graph construction plays an important role on the efficiency of algorithms. Low-rank representation is a powerful tool for exploring the multiple subspace structures of data. It can capture the global structure of the data and is robust to noises, thus it has been widely applied in many areas. To address the problem of graph construction in feature extraction algorithms with low-rank and sparse decomposition technique, this project proposes to research: (1)constructing low-rank and sparse graphs based on existing low-rank representation methods, and then designing novel feature extraction algorithms with constructed graphs based on classifical algorithms, (2)a framework of simultaneous graph learning and feature extraction, which can utilize data information in different spaces, (3)making use of characteristics and prior information of the data to guide the construction of low-rank representation models which are helpful for feature extraction. Compared with traditional feature extraction algorithms, the research cannot only handle and analysis high-dimensional data more effectively, but also can improve the robustness to noises, the efficiency of parameter selection and feature extraction. Moreover, the research provides theoretical support for the related research in computer vision and pattern recognition areas.

英文关键词： subspace method;discriminant subspace;projection;linear discriminant analysis;discriminant vector