Modified Particle Swarm Optimization with Adaptive Inertia, Constriction Coefficient, and Lévy-Flight Perturbation for Optimal Feature Selection in Face Recognition

Abstract

Face recognition (FR) is one of the fundamental challenge faced in computer vision with it’s applications spanning from security, surveillance to human–computer interaction. The high dimensionality of facial feature spaces extracted via Principal Component Analysis (PCA) introduces the curse of dimensionality, motivating effective feature selection. This paper proposes a Modified Particle Swarm Optimization (MPSO) algorithm that integrates three synergistic enhancements into binary PSO: (i) an adaptive linearly decreasing inertia weight for progressive exploration–to–exploitation transition; (ii) the Clerc–Kennedy constriction coefficient for theoretically guaranteed velocity convergence; and (iii) a Lévy-flight perturbation mechanism applied stochastically to personal best positions to escape local optima.Binary feature masks are produced via a V-shaped transfer function for symmetric bit-flip behaviour. The selected PCA features drive a k-Nearest Neighbour (k-NN) classifier for identity verification. Experiments conducted on the ORL and Extended Yale-B databases demonstrate that MPSO achieves 97.25% and 91.83% recognition accuracy, respectively, which is within 1.85% of VGGFace2-CNN on ORL while training exclusively on the target dataset with no GPU requirement. It outperforms standard PSO, GA, DE, GWO, and WOA by selecting fewer than 40% of available PCA features. The total computational overhead of MPSO over standard PSO is only 5.6% (47.3s vs. 44.8s per run). Statistical significance is confirmed through Wilcoxon rank-sum testing, Cohen’s d effect sizes, Friedman’s test, and Holm-Bonferroni post-hoc correction test. MPSO thus provides a strong, efficient, and interpretable solution for face recognition in constrained environments where deep learning deployment is infeasible.