一种基于指数降维的监督型稀疏保持典型相关分析算法

An improved supervised sparsity preserving CCA algorithm based on exponential dimensionality reduction

提出一种基于指数降维的监督型稀疏保持典型相关分析算法.通过将样本的类别信息与样本特征相融合,克服以往引入监督信息导致重建误差增大的缺陷,同时实现类内相关的最大化与类间相关的最小化;针对传统算法处理稀疏信号的高维小样本问题的瓶颈,改进算法对总体散布矩阵做指数化的处理,既保留有效信息,又将总体散布矩阵非奇异化,克服PCA预处理散布矩阵导致有效信息流失的缺陷.依据ORL,Yale,AR和FERET人脸数据库而进行的仿真实验表明,该算法比其他的典型相关分析方法具有更好的识别效果.

An improved supervised sparsity preserving canonical correlation analysis algorithm based on exponential dimensionality reduction was proposed. The problem that the fitting error increased while adding supervised information to the SPCCA was solved by the fusion of the class label information and sample feature. The local manifold structure of the data was realized at the same time. Aimed at the problem of traditional algorithm in dealing with small sample of high-dimensiona sparse signal,index scattering matrix was used to retain effective information while building the non-singular scattering matrix. It overcame the default of effective information losses while using PCA to extract principal features of the scattering matrix.The experimental results on ORL,Yale,AR and FERET face databases showed that the proposed algorithm was better than related canonical correlation analysis methods in recognition effect.