2008 年 37 巻 3 号 p. 231-243
A multipose human face recognition approach is presented. The proposed scheme is based on frequency analysis (i.e. DCT or wavelet transforms) to obtain facial features which represent global information of face image and modified LDA (M-LDA) to classify the facial features to the person's class. The facial features are built by selecting a small number of frequency domain coefficients that have large magnitude values. Next, from the facial features, the mean of each face class and the global covariance are determined. Finally, by assuming that each class has multivariate normal distribution and all classes have the same covariance matrix, M-LDA is used to classify the facial features to the person's class. The aims of proposed system are to reduce the high memory space requirement and to overcome retraining problem of classical LDA and PCA. The system is tested using several face databases and the experimental results are compared to well-known classical PCA, LDA, and other established LDA (i.e. DLDA, RLDA, and SLDA).