基于Kernel Discriminant Canonical Correlation(KDCC)的多观测样本分类算法

The Classification Algorithm of Multiple Observation Samples Based on Kernel Discriminant Canonical Correlation

针对多观测样本分类问题,提出一种基于Kernel Discriminant CanonicalCorrelation(KDCC)来实现多观测样本分类的模型.该算法首先把原空间样本非线性的投影到高维特征空间,通过KPCA得到核子空间,然后在高维特征空间定义一个使类内核子空间的相关性最大,同时使类间核子空间的相关性最小的KDCC矩阵,通过迭代法训练出最优的KDCC矩阵,把每个核子空间投影到KDCC矩阵上得到转换核子空间,采用典型相关性作为转换核子空间之间的相似性度量,并采用最近邻准则作为多观测样本的分类决策,从而实现多观测样本的分类.在三个数据库上进行了一系列实验,实验结果表明提出的方法对于多观测样本分类具有可行性和有效性.

A novel pattern recognition model, which is the classification algorithm of multiple observation samples based on kernel discriminant canonical correlation is proposed in this paper. The original input space is nonlinear mapping into a high-dimensional feature space, and each image set is represented by a kernel subspace using KPCA. Then define a KDCC matrix, which maximizes the similarities of within-subspaces, and simultaneously minimizes those of between-kernel subspaces. The paper proposes an iterative algorithm to train the optimal KDCC matrix, and each kernel subspace projects into the KDCC matrix so as to obtain the transformed kernel subspace. Use canonical correlation to solve the similarity between the transformed kernel subspaces. Then the nearest neighbor classifier is used to solve the classification of multiple observation samples. Experiments on three types of databases prove that the proposed method is valid and efficient.