Tackling Challenge of Masked and Unmasked Face for Facial Recognition Through Dimensionality Reduction and Deep Learning

Abstract

Facial recognition systems face considerable performance degradation in the presence of occlusions, particularly face masks, which are increasingly common in public setting. This research examines the impact of such occlusions by employing dimensionality reduction techniques: principal component analysis (PCA) and auto-encoders, combined with convolutional neural networks for classification. Through experiments on custom-built dataset comprising masked, unmasked, and mixed facial images, we observe that PCA consistently outperforms auto-encoders across all evaluation metrics, including accuracy, precision, recall and F1-score. However, PCA is more sensitive to occlusion variability, while auto-encoders exhibit more stable, though generally lower performance. Notably, performance declines significantly when models are evaluated on mixed datasets, emphasizing the complexity of realworld scenarios. Data augmentation yields only marginal improvement, underscoring the limitation of current approaches in handling diverse occlusions. The study also explores the potential of advanced mathematical tools such as higher-order tensor decomposition and fractional differentiation to enhance feature representation and model robustness. Overall, this research directions involving diverse training sets, sophisticated feature extraction and hybrid architectures to enhance model generalization in dynamic environments. Hence, contributing to the development of more reliable biometric systems.