Feature selection using order statistics

One of the fundamental problems in statistical pattern recognition, particularly in face recognition and similar applications, is the intractably high number of dimensions used to represent the input data. Various dimensionality reduction techniques have been studied in recent literature. Many of these techniques, including Principle Component Analysis (PCA), can be divided into two parts: feature extraction and feature selection. Traditional methods for feature selection have been either naive or computationally expensive. In this report, a new noise-resistant method for feature selection based on order statistics is proposed. Experimental results for the selection of PCA features in face recognition show that the new feature selection algorithm gives superior or comparable performance compared to the traditional naive feature selection method in the presence of noise, especially when the number of classes is small compared to the number of training examples per class.