Abstract: In this paper，based on the Canonical Correlation Analysis（CCA），a new supervised learning algorithm called Locality Discriminative CCA（LDCCA） is developed，which introduces the class information of samples and considers the local correlations of both of the within-class samples and the between-class samples.The feature extracted by LDCCA can realize the maximization of local within-class correlations and the minimization of local between-class correlations，which is good for classification of pattern.The experimental results on an artificial dataset，multiple feature database and facial database including ORL，Yale and AR show that the proposed LDCCA can effectively enhance the classification performance by using class information.

Key words: canonical correlation analysis, feature extraction, local discrimination, dimensionality reduction

摘要： 在典型相关分析（CCA）的基础上，通过引入样本的类信息，并结合局部化思想，充分考虑了同类样本之间的局部相关与不同类样本之间的局部相关关系及其对分类的影响，提出了一种新的有监督学习方法——局部判别型CCA（Locality Discriminative CCA，简记为LDCCA）。LDCCA提取的特征能够实现同类样本之间相关最大化，同时使得不同类样本之间相关最小化，这将有利于模式的分类。在人工数据集，手写体数字数据集上和ORL，Yale和AR人脸数据集的实验结果表明，LDCCA能有效地利用类信息来提高分类性能。

关键词: 典型相关分析, 特征提取, 局部化判别, 降维