A Lightweight Transfer Learning Framework for Masked Face Recognition Using Feature Extraction with Ensemble Decision Fusion and Cross-Validation

Masked face recognition remains a major challenge in biometric authentication due to occlusion, illumination variation, disguises, and diverse environments. We propose RACE (RetinaFace detection with ArcFace embeddings and Classifier Ensemble) approach, a lightweight yet robust framework integrating face alignment, discriminative feature extraction, and decision-level fusion. RetinaFace performs detection and alignment, ArcFace generates compact embeddings, and three complementary classifiers—Prototype, k-Nearest Neighbors (kNN), and Linear head—are combined through majority voting. The framework was rigorously evaluated on three benchmark datasets (AR, IIIT-Delhi Disguise Face v1, and MFD) under challenging scenarios including masks, scarves, sunglasses, heavy disguises, pose variation, and illumination changes. Across six experimental settings, RACE achieved near-perfect accuracy (mostly 100%, others ≥99.26%), with AUC values approaching 1.0. Generalization was confirmed using stratified 5-fold cross-validation with Wilson-pooled 95% confidence intervals, yielding accuracies between 96.94% and 100% across all experiments. Statistical analyses like Chi-square and Cochran’s Q tests further validated classifier significance. To enhance interpretability, we applied the Local Interpretable Model-Agnostic Explanations (LIME) framework to each classifier, identifying key feature contributions in individual predictions. Overall, RACE provides a highly reliable solution for robust masked face recognition in real-world conditions.