Abstract: Class discriminant ability of the feature and the feature space is put forward to analyze the performance of the feature spaces given by PCA，ICA and LDA，and then diagram form is drawn to compare the discriminant ability of the above three methods.Genetic algorithm is improved to avoid “over-fitting”.At last，a new face feature extraction method which is high order independent and can evaluate the class information is brought forward：ICA is used to extract high order independent features which are used as the original solution set and improved GA has been adopted to combine class information.Experimental results show that the scheme can obtain good results.

Key words: principal component analysis, independent component analysis, linear discriminant analysis, genetic algorithm

摘要： 提出了特征及特征空间的类别鉴别能力的概念，并对PCA、ICA、LDA提取的特征空间的鉴别能力作了评价；然后采用ICA提取的特征作为初始解集，以特征空间的鉴别能力作为评价准则，用遗传算法来进行人脸特征选择；为了避免“过训练”现象，对遗传算法进行了改进。采用该特征提取方法，可以得到既高阶独立或近似独立又可以达到类内差异最小、类间差异最大的特征子空间。实验结果表明，该特征提取方法可以取得良好效果。

关键词: 主成分分析, 独立分量分析, 线性判别分析, 遗传算法