项目名称: 矩阵对齐的耦合距离度量学习方法研究
项目编号: No.61201370
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 电子学与信息系统
项目作者: 贲晛烨
作者单位: 山东大学
项目金额: 24万元
中文摘要: 传统的距离度量学习仅能对同维数的向量和同尺寸的图像矩阵等单一数据集合进行度量,而对来自不同集合的数据度量是没有意义的。本项目针对此问题,同时为了避免将矩阵样本转成向量对原始特征结构信息造成的破坏,将基于矩阵的子空间降维方法的思想引入到向量耦合距离度量学习之中,对矩阵对齐的耦合距离度量学习理论及其解耦问题进行研究。本项目将针对不同集合下的矩阵样本建立合理联系和匹配关系问题,提出矩阵耦合距离度量的概念;将边信息、局部保留、相关关系和类别等监督信息引入到矩阵对齐的耦合距离度量学习理论之中,提出耦合距离度量学习的新算法;并建立其统一的度量学习框架;寻找其与二维典型相关分析算法的关系。本项目所研究的方法将成为解决不同分辨率图像的匹配问题和同一个体但不同模态图像匹配问题的重要手段,研究成果易推广到各种图像的匹配上,为相关工程实践提供重要的技术支持和理论保证。
中文关键词: 矩阵对齐;耦合距离度量学习;局部保留;类别信息;联立类别信息
英文摘要: Traditional distance metric learning can measure only single data set like vectors of same dimension or image matrices of identical size. But the flaw lies in the fact that it has nothing to do with the measurement of different data sets. What's more, destroying of original structure information happens when 2D image matrix is transformed into 1D vector. Thus, aiming at the solution to these two problems, research will be conducted on coupled distance metric learning with matrix alignment and its decoupling problem by introducing the idea of matrix-based subspace dimensional reduction. In order to provide reasonable connection and matching relation among different kinds of collections with material effective algorithms,the concept of matrix coupled distance metric will be proposed; By introducing some supervised information, such as side information, locality preserving, correlation and discriminative information, to coupled distance metric learning with matrix alignment, novel distance metric learning algorithms will be put forward and the unified metric learning frame will be estabilished. Furthermore, the relationship between the proposed theroy and Two dimensional canonical correlation analysis (2DCCA) will be searched. Supported by this project, the proposed methods are to be important solutions to the prob
英文关键词: matrix alignment;coupled distance metric learning ;locality preserving;discriminative information ;linked discriminative information