摘要
核Fisher鉴别分析(KFDA)已成为抽取非线性特征的最有效方法之一.针对在解决两类模式分类问题中KFDA只能获得一个鉴别矢量的弱点,提出了一种改进的核Fisher鉴别分析(MKFDA)方法,该方法对特征空间中的两类间离散度进行了重新估计,通过使用核类间散布矩阵的一种特殊形式,我们可以得到最多N(N为训练样本数)个鉴别矢量,从而提高了两类模式问题的分类性能.在IRIS数据上的实验结果验证了MKFDA方法的有效性.
Although kernel-based Fisher discriminant analysis （KFDA） has became one of the most effective techniques for nonlinear feature abstraction, it can only obtain one discriminant vector when solving the problem of classification between two class. Aiming at this deficiency of KFDA, a modified kernel-based Fisher discriminant analysis（MKFDA） is put forward, in which the between-class scatter degree of two class is re-estimated in the feature space, by using this special form of kernel-based between-class scatter matrix, at most N （the number of training sample） discriminant vectors can be obtained, so the classification ability can be improved largely. Test result on IRIS data set shows the validity of the MKFDA.
出处
《淮阴师范学院学报（自然科学版）》
CAS
2005年第4期335-340,共6页
Journal of Huaiyin Teachers College;Natural Science Edition
基金
江苏省教育厅自然科学基金资助项目(04KJD520037)
关键词
核FISHER鉴别分析
核技巧
类间散布量
特征抽取
非线性最佳鉴别特征
kernel-based Fisher discriminant analysis
kernel trick
between-class scatter degree
feature extraction
nonlinear optimal discriminant feature
