Enhancing Facial Recognition Performance with Data Augmentation in Occluded Environments

Facial recognition systems frequently face challenges in accurately identifying individuals when critical facial features are obscured by occlusions such as masks, sunglasses, or scarves. These scenarios degrade the reliability of recognition models, especially in real-world environments. Data augmentation has emerged as a powerful strategy to improve model generalisation by artificially increasing dataset diversity and simulating occlusion-rich conditions. This study investigates the role of augmentation techniques, which include rotation, mirroring, zooming, and brightness adjustment, in enhancing recognition accuracy under partial and full occlusions. Experimental evaluation demonstrates that models trained with data augmentation achieve notable improvements over non-augmented baselines. For instance, average recognition accuracy improved from 72.4% to 86.3% under face mask occlusions, from 70.1% to 84.7% under sunglasses occlusions, and from 68.9% to 82.5% under scarf occlusions. When augmentation was combined with illumination normalisation, further gains were observed, with overall accuracy reaching 88.9% and F1-scores exceeding 87%. These results confirm that data augmentation substantially improves resilience against occlusions, while combined augmentation pipelines provide additional robustness in variable lighting and pose conditions. The findings highlight data augmentation as a foundational strategy for developing more resilient facial recognition systems. This work advances recognition performance in occlusion-heavy environments, with implications for applications in security, surveillance, and identity verification.