Abstract: A general and effective method is put forward to design the RBF neural network classification, which is used to recognize face in the paper. In order to avoid a fitting and reduce computation cost, the Fisher and linear discriminative technology is used to reduce dimension and extract face feature. The mixture learning algorithm is used to train RBF neural network, and so the gradient descent method is greatly reduced in search space. A new information clustering algorithm is introduced to train sample classification, and so all of the same data can be gathered together. At the same time, the compact structure RBF neural network classifier is also used to improve the algorithm. At last, the simulation experiments are done based on ORL database, and the results show that the algorithm achieves high efficiency and effectiveness.

Key words: Radial Basis Function（RBF）, linear classification, linear discriminant function, clustering algorithm

摘要： 提出了一个通用而且有效的方法来设计RBF神经网络分类器用于人脸识别。为了避免过拟合和减少计算量，用主元分析法和Fisher线性判别技术来降低维数，以提取人脸特征；利用一个混合的学习算法来训练RBF神经网络，使梯度下降法的搜索空间大大减少；采用一种基于训练样本类别信息的新的聚类算法，所有同类的数据可被聚集在一起，尽量减少不同类数据混杂在一起，同时选取结构尽可能紧凑的RBF神经网络分类器。在ORL数据库上进行了仿真，实验结果表明，该算法具有高效性和有效性。

关键词: 径向基, 线性分类, 线性判别式, 聚类算法