关键词: 统计不相关特征, 特征抽取, 无监督模式, 人脸识别

Abstract: Uncorrelated optimal discriminant plane is an important feature extraction method and has been wildly used in the pattern recognition field. However, uncorrelated optimal discriminant plane is based on Fisher criterion function and the conjugated orthogonal constraint of the totalclass scatter matrix. It needs the class information to calculate the Fisher optimal discriminant vector. Thus, it can only be used in supervised pattern. A new method was presented to extend uncorrelated optimal discriminant plane to unsupervised pattern. The basic idea was to introduce the fuzzy concept into Fisher linear discriminant analysis and used the defined fuzzy Fisher criterion function as its optimized objective to figure out an optimal discriminant vector and fuzzy scatter matrixes in unsupervised pattern. In the conjugated orthogonal constraint of the fuzzy totalclass scatter matrix, the second discriminant vector which maximized the fuzzy Fisher criterion can be obtained. Thus, a new feature extraction method based on unsupervised uncorrelated optimal discriminant plane was proposed. The experimental results for UCI datasets and CMU-PIE face database demonstrate that although this method is unable to surpass traditional uncorrelated optimal discriminant plane, it can extract the uncorrelated features for classification and is superior to the common unsupervised feature extraction methods such as principal component analysis and independent component analysis when the between-class difference is big.

Key words: Statistical uncorrelated features, feature extraction, unsupervised pattern, face recognition