Robust Face Recognition in the Presence of Diverse challenges: A Hybrid Deep Neural Network Approach

This paper presents a novel lightweight hybrid architecture for face recognition, combining the strengths of MobileNet and attention mechanisms to enhance performance under challenging conditions such as facial occlusions (e.g., masks), varied illumination, and diverse expressions. The proposed model is evaluated against popular baseline models, including MobileNetV2, EfficientNetB2, and VGG16, on the Yale Face Dataset and a Simulated Masked Yale Dataset. On the Yale Dataset, the hybrid model achieved superior results with an accuracy of 93.78%, precision of 94.45%, recall of 93.33%, and F1-score of 93.89%, outperforming the baseline models in all key metrics. Additionally, when tested on the Simulated Masked Yale Dataset, the hybrid model exhibited increased resilience to occlusion with an accuracy of 63.45% and F1-score of 64.22%, significantly surpassing the other architectures.