基于样本分布相似度的期望分布鉴别分析

Expected Distribution Discriminant Analysis Based on Similarity of Sample Distribution

主分量分析(PCA)和线性鉴别分析(LDA)是模式识别领域使用广泛的两种特征抽取方法.本文针对两种方法的不足之处,并从样本分布相似度出发提出一种期望分布鉴别分析(EDDA)方法,抽取到的鉴别特征的总体分布和设定的期望分布最为相近.即通过 EDDA 得到的投影向量可以抽取出最接近理想分布的鉴别特征.EDDA 在投影向量的求解问题上不存在小样本问题,抽取的鉴别特征维数小,并且整体识别性能得到增强.在 ORL、Yale 人脸库上的实验结果证明本文方法在人脸识别精度上优于 PCA 和 LDA 方法.

Principal component analysis （PCA） and linear discriminant analysis （LDA） are two kinds of popular feature extraction methods for pattern recognition. A new method, expected distribution discriminant analysis （EDDA）, is proposed based on the similarity of sample distribution after some disadvantages of PCA and LDA are indicated. The distribution of extracted features is mostly close to the expected distribution such as idealized distribution by using EDDA. Based on EDDA, the small sample size problem （SSSP） does not occur any more. The dimension of discrimination feature is very low and the recognition performance is enhanced. Some experimental results on ORL and Yale face database demonstrate that the proposed method has higher recognition rate than PCA and LDA.