关键词: 核鉴别分析, 非线性特征抽取, 新的核类间散布矩阵, 最小距离分类器

Abstract: Kernel-based nonlinear Discriminant Analysis（KDA） has attracted much attention due to the high performance and kernel-based nonlinear discriminant algorithms are developed. Among these algorithms, one of its major disadvantages is that the maximal number of its discriminating vectors capable to be found is limited by the number of classes involved, i.e., L－1 for L-class problem. For binary-class problem, it proposes a new KDA to break through the inherent limitation by a special form of kernel between-class scatter-matrix. Experimental results show that the approach gives impressive recognition performances compared to both the Alternative Fisher Linear Discriminant Analysis（AFLDA） and the Fisher Linear Discriminant Analysis（FLDA）.

Key words: kernel-based discriminant analysis, nonlinear feature extraction, new kernel between-class scatter matrix, minimum distance classifier