Significance Test for Feature Subset Selection on Image Recognition

Abstract

This paper proposes a novel feature selection method based on significance test (ST). Statistical significant difference between (or among) classes, such as t statistic in Student test and F statistic in ANOVA, is utilized to measure pattern recognition ability of individual features. The feature significance level during a feature selecting procedure is used as feature selection criterion, which is determined by the product of the significant difference level and the independent coefficient of the candidate feature. An algorithm of maximum significant difference and independence (MSDI) and strategies of monotonically increasing curve (MIC) are proposed to sequentially rank the feature significance and determine the feature subset with minimum feature number and maximum recognition rate. Very good performances have been obtained when applying this method on handwritten digital recognition data.