Deep Learning Architectural Enhancements for Robust Facial Recognition in Occluded Scenarios

Abstract

Background: Facial recognition systems often experience reduced performance when key facial regions are obscured by masks, sunglasses, scarves or hand overlays. Aims: The study aims to investigate architectural enhancements to standard Convolutional Neural Network (CNN) models, namely Residual Network (ResNet), Mobile Neural Network (MobileNet) and Inception Network, to improve facial recognition accuracy under occluded conditions such as masks, sunglasses and scarves. Study Design: An experimental comparative study was conducted to evaluate baseline CNN architectures against architecturally enhanced versions incorporating attention mechanisms, feature fusion layers, dropout and batch normalisation. Place of Study: Southern Delta University, Ozoro, Delta State, Nigeria. Methodology: A combined dataset of approximately 25,000 facial images was assembled from benchmark sources, including Labeled Faces in the Wild (LFW) (Huang et al., 2007), the AR face dataset (Martinez & Benavente, 1998) and the CelebFaces Attributes (CelebA) dataset (Liu et al., 2015), supplemented with synthetic occlusion variants (masks, sunglasses, scarves and hand overlays) generated through data augmentation. ResNet-50, MobileNet and Inception backbones were enhanced with self-attention and spatial attention layers and multi-scale feature fusion modules, then trained using the Adam optimiser (learning rate 0.001, batch size 64, up to 120 epochs and early stopping after 10 stagnant epochs), with dropout (0.3-0.5) and batch normalisation applied for regularisation. Performance was assessed using precision, recall, F1-score, False Acceptance Rate (FAR) and False Rejection Rate (FRR), comparing enhanced models against their baseline counterparts. Results: The enhanced ResNet attained 88.4% precision, 86.7% recall and an 87.5% F1-score, compared with 74.2%, 71.8% and 73.0%, respectively, for the baseline ResNet. MobileNet with feature fusion reached an F1-score of 85.3%, compared with 70.6% for baseline MobileNet. The enhanced Inception model achieved 89.1% precision, 87.4% recall and an 88.2% F1-score. Error analysis showed that the enhanced ResNet reduced FAR to 3.6% from 7.9%, while MobileNet with feature fusion reduced FRR to 5.1% from 10.2%. Overall, enhanced models improved precision, recall and F1-score by 8-12% across occlusion scenarios relative to baseline CNNs. Conclusion: Integrating attention mechanisms and feature fusion layers with training optimisations such as dropout and batch normalisation substantially strengthens the robustness of CNN-based facial recognition systems under occlusion. These architectural enhancements show potential for biometric authentication and controlled security applications. Deployment in sensitive domains, including law enforcement, should be preceded by comprehensive fairness evaluation, privacy safeguards, legal compliance and human oversight, although further work is needed to assess computational efficiency and real-time adaptability.