目的 针对2维线性鉴别分析提取人脸特征向量稳定性较差、仅对行或列方向提取特征时容易丢失不同行或列间有助于鉴别分析的协方差信息、同时存在特征维数较高的问题，提出一种广义并行2维复判别分析的人脸识别方法。方法 首先对人脸图像进行广义并行2维线性判别分析处理，根据特征值贡献率动态选取特征向量组成正交投影矩阵，完成水平和垂直方向上的投影；其次将处理后得到的两类特征矩阵以复数的实部和虚部形式相加，对融合后的特征矩阵进行广义2维复判别分析处理得到复特征矩阵；然后以复特征矩阵的特征值大小来衡量特征矩阵分量的识别性能，对特征矩阵分量进行重新排序，选取最具鉴别力的分量形成最终表征人脸的特征；最后采用最大相似度分类器比较测试样本与训练样本特征的相似度，进行人脸图像特征的分类识别。结果 在Yale、ORL、FERET、CMU-PIE及LFW人脸数据库上进行实验测试，该方法的最优识别率分别为100%、100%、98.98%、99.76%及98.67%，特征维数在8590之间，表明该方法对复杂条件下的人脸识别有较高的准确率和较低的空间占有率。结论 该方法能够有效克服2维线性鉴别分析提取特征稳定性差、特征空间中特征重叠、存储系数多、特征维数高的缺点，表现出较高鲁棒性和准确率及较低空间复杂度的特性。

Objective A face recognition approach of generalized parallel two-dimensional (2D) complex discriminant analysis was proposed to tackle such problems that 2D linear discriminant analysis demonstrated poor stability when extracting facial feature vectors, the covariance information of different rows or columns which was conducive to discriminant analysis was very likely to get lost when only features in rows or columns were being extracted, and the dimensions where features existed were relatively high. Method Firstly, generalized parallel 2D linear discriminant analysis was conducted on facial images, and the feature vectors are selected according to the feature value contribution rate to form the orthogonal projection matrix, then the projection of horizontal and vertical direction is completed; secondly, the two types of feature matrices obtained after processing were added together in forms of real part and imaginary part of complex numbers, and the complex feature matrices were obtained by conducting generalized 2D complex discriminant analysis on feature matrices having been fused; then, the recognition performance of feature matrix components was measured based on feature values of complex feature matrices, the feature matrix components were re-ranked, and the most discriminative components were selected to form the final features characterizing human faces; and at last, maximum similarity classifier was used to classify and recognize features of human face images by comparing the similarity between the test samples and the training sample features. Result Yale, ORL, FERET, CMU-PIE and LFW face databases were experimented, from which the optimal recognition rates obtained by using this method were respectively 100%, 100%, 98.98%, 99.76%, and 98.67%, with the feature dimensions ranging from 85 to 90, which indicated that this method delivered relatively high face recognition precision and low space occupancy in complex conditions. Conclusion This method could effectively overcome drawbacks such as poor feature extraction stability of 2D linear discriminant analysis, overlap of features in feature space, excessive storage coefficients, and high dimension of features, manifesting high robustness, great precision, and low space complexity.