Abstract: Aiming at the problem that data dimensionality reduction methods are often difficult to balance the global and local feature information in the feature extraction stage of face recognition, and Bayesian classifier exists the small sample problem in the matching recognition stage, a Bayesian face recognition method based on global and local feature fusion is proposed. In this method, the global non-linear feature of face data is extracted by kernel principal component analysis, and on this basis, the local manifold structure information of face data is extracted by the orthogonal locality sensitive discriminant analysis, achieving the purpose of extracting the low-dimension essential facial feature with high-discrimination. Then, a maximum entropy covariance selection method is utilized to estimate the covariance matrix to solve the small sample problem in Bayesian classifier design. Experiments are designed on ORL, AR, YALE and FLW face database. The results show that the proposed feature extraction algorithm and the improvement of Bayesian classifier have achieved good results, and the optimization of these two stages can improve the effect of face recognition significantly.

Key words: face recognition, global feature, local feature, feature fusion, Bayesian classifier

摘要： 针对人脸识别特征提取阶段中的数据降维方法往往难以兼顾保持全局与局部特征信息的问题，以及匹配识别阶段贝叶斯分类器中小样本问题，提出了一种融合全局与局部特征的贝叶斯人脸识别方法。该方法通过核主元分析提取出人脸数据的全局非线性特征，并在此基础上通过正交化局部敏感判别分析挖掘出人脸数据的局部流形结构信息，以达到提取出具有高判别力低维本质人脸特征的目的；采用一种最大信息量协方差选择的方法，来对协方差矩阵进行估算，以解决贝叶斯分类器设计中的小样本问题。在ORL、AR、 YALE、FLW人脸库上设计实验来进行验证。结果表明，提出的特征提取算法以及对贝叶斯分类器的改进取得了比较好的效果，通过对这两个阶段的优化，可以显著提升人脸识别的效果。

关键词: 人脸识别, 全局特征, 局部特征, 特征融合, 贝叶斯分类器