Improve Handwritten Character Recognition Performance by Heteroscedastic Linear Discriminant Analysis

This paper presents a novel LDA algorithm named 2DHLDA (2-dimensional heteroscedastic linear discriminant analysis). The proposed algorithms are applied on age-group classification using facial images under various lighting conditions. 2DHLDA significantly overcomes the singularity problem, so-called ´small sample size´ problem (S3 problem), and the original feature space is split into useful dimensions and nuisance dimensions to reduce the influence of different lighting conditions. A two-phased dimensional reduction step, namely 2DHLDA+LDA, is used in our experiment. Our experimental results show that the new 2DHLDA-based approach improves classification accuracy more than the conventional 1D and 2D-based approaches