稀疏保持典型相关分析特征选择与模式识别  

Feature Extraction and Pattern Recognition Based on Sparsity Preserving Canonical Correlation Analysis

作　　者：许洁[1,2] 吴秦[1,2] 梁久祯[3] 王念兵[1,2] 张淮[1,2] 

机构地区：[1]江南大学物联网工程学院智能系统与网络计算研究所,江苏无锡214122 [2]物联网技术应用教育部工程研究中心,江苏无锡214122 [3]常州大学信息与工程学院,江苏常州213164

出　　处：《小型微型计算机系统》2017年第8期1877-1882,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61202312;61673193)资助;中央高校基本科研业务费专项资金(USRP51635B;JUSRP51510)资助;江苏省自然科学基金项目(BK20150159)资助

摘　　要：构建一种基于稀疏保持典型相关分析(Sparsity Preserving Canonical Correlation Analysis,SPCCA)的特征提取算法.在典型相关分析(Canonical Correlation Analysis,CCA)的基础上,利用稀疏保持投影(Sparsity Preserving Projections,SPP)将样本之间的类别信息引入,从而提高识别率.首先,提取同一类训练样本的两组特征向量,由SPP构建稀疏重构权重矩阵;并建立相应的判据准则函数来描述两组特征向量之间的关系;然后根据所建立的准则函数提取出相应的典型投影矢量集;最后将两组特征通过不同的特征融合策略进行融合,并在分类识别中进行应用.典型相关分析算法将两组特征向量的相关性特征作为有效信息来用于分类,既能融合信息,又能过滤冗余信息.同时SPCCA算法中引入了两组特征的权重矩阵,在一定程度上提高了融合特征的鉴别力,提高了识别精度.在MFEAT、UCI手写体数字库及ORL、PIE人脸数据库上的实验结果证明了本文方法的有效性和稳定性.A feature extracting algorithm based on Sparsity Preserving Canonical Correlation Analysis （ SPCCA ） was constructed in this paper. Based on Canonical Correlation Analysis （CCA） ,the class information of samples was introduced by using sparsity preserving projection and the recognition rate was siginificanfly improved. First, extracting two groups of feature vectors with the same training samples, and constructing sparsity reconstruction weight matrix by sparse preserving projection;second, establishing the correlation criterion function between the two groups of feature vectors;third, extracting two groups of canonical projective vectors according to this function; four,doing feature fusion for classification by using different strategy. CCA algorithm takes the correlation feature between the two sets of feature vectors as the effective information, which suits for information fusion and eliminates the redundant informarion within the features. And the weight matrix of two sets of features were introduced in SPCCA algorithm, which enhances the ability of identify the characteristics of the fusion feature and improves the recognition accuracy. The results of experiments on MFEAT, UCI handwritten digit database and ORL, PIE face database show that this algorithm is efficient and robust.

关 键 词：稀疏保持投影 典型相关分析 特征融合 组合特征提取 人脸识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]