基于无监督最佳鉴别平面的人脸识别

Face recognition based on unsupervised optimal discriminant plane

最佳鉴别平面作为一种重要的特征抽取方法,在人脸特征降维中具有重要的影响。然而,传统的最佳鉴别平面是基于Fisher准则的,只能用于有监督模式。为此,提出了一种将最佳鉴别平面扩展到无监督模式下的方法,其基本思想是以投影空间中模糊类间离散度和模糊类内离散度的比值最大为优化目标,计算出无监督模式下最佳鉴别矢量及模糊离散度矩阵,进而获得一种新的基于无监督最佳鉴别平面的特征抽取方法。较之同属于无监督特征抽取的主成分分析,该方法更容易获得有利于分类的特征。对CMU-PIE人脸数据库进行实验,结果表明,在样本类别信息缺失的情况下,该方法尽管无法具有与有监督模式下的最佳鉴别平面特征抽取方法同样的性能,但当不同类之间差异较大时,将优于主成分分析方法。

As an important supervised feature extraction method, optimal discriminant plane has great influence in the facial feature reduction. However, traditional optimal discriminant plane is based on Fisher criterion function and it can only be used in supervised pattern. This paper presented a novel method to extend optimal discriminant plane to unsupervised pattern. Using the maximum ratio between fuzzy between-class scatter and within-class scatter in the projection space as its optimization objective, an optimal discriminant vector and fuzzy scatter matrixes in unsupervised pattern could be figured out. With these, obtained a novel feature extraction method based on unsupervised optimal discriminant plane. Compared to principal component analysis（PCA） unsupervised feature extraction algorithm, this method is easier to get the features for the classification. The experimental results for CMU-PIE face database demonstrate that although this method can’t have the same performance of traditional optimal discriminant plane in the condition of unlabelled data, it is superior to principal component analysis when the difference between one class and the another class is big.