关键词: 人脸识别, 线性辨别分析, 特征分解, 辨别能量, 聚类分析

Abstract: Recently， current researches indicated that， the angle between the eigenvector corresponding to the largest eigenvalue of the inter-class covariance and the eigenvector corresponding to the largest eigenvalue of the intra-class covariance is more crucial to the performance of traditional linear discriminant methods， furthermore， if the two eigenvectors are parallel； the final results may be disputable. However， upon careful scrutiny on his assertion， conclude that the angle between the two eigenvectors is less decisive to the performance， more over； the main drawback of traditional linear methods is the inter-class covariance cannot precisely reflect the discriminant information. Simply maximizing the inter-class covariance in the principle component space may induce the losing of adjacent class-pair’s contribution. Therefore， propose the optimal linear discriminant analysis （OLDA） method， which distributes equivalent authority for each class-pair by employing “discriminative power”. Besides， employ the gradient scheme to derive the feature vectors， and a constraint condition is introduced to evaluate the convergence speed. Thirdly， to address the multimodal problem， the pre-clustering mechanism is adopted to ameliorate the nonlinear structure. Apply this method on a practical face database and a virtual database， the experimental results show the promise of this method.

Key words: face recognition, linear discriminant analysis, eigen-decomposition, discriminant power, clustering