Linear Discriminant Analysis (LDA) for Ear Recognition: Methods Comparison and Implementation

Linear Discriminant Analysis (LDA) is a extensively utilized linear classification method for feature extraction, but its direct application to ear recognition encounters dimensionality issues and small sample size problems. To overcome these challenges, researchers have explored multiple pathways and successfully implemented LDA-based ear recognition systems. This paper conducts theoretical comparisons of several LDA-based ear recognition methods with supporting experimental data, including Fisherears, DLDA, VDLDA, and VDFLDA. The implementation typically involves preprocessing ear images, extracting feature vectors using covariance matrix computations (through MATLAB's cov function or Python's numpy.cov), and applying eigenvalue decomposition for dimensionality reduction. Experimental results demonstrate that VDFLDA, which incorporates variance normalization and flexible discriminant analysis, outperforms other methods by achieving better separation between classes while handling limited training samples efficiently. By implementing VDFLDA, ear recognition accuracy and performance can be significantly enhanced through optimized scatter matrix calculations and improved projection vector derivation.