Revealing the Unseen: Explainable AI-Driven Masked Face Recognition
Madhu Kashania, Ankit Rajpal, Sheetal Rajpal, Naveen Kumar · 2024
Face recognition provides a biometric authentication mechanism that has evolved significantly with the advancement of deep learning. However, the emergence of COVID-19 mandates wearing face masks. Since then, people have been more cautious about wearing face masks as a preventive measure for air-transmitted diseases, flu, and pollution. This poses a challenge for face recognition systems because a significant portion of the face is covered with a mask. To solve the problem of masked face recognition, several machine learning and deep learning models, including attention modules, have been employed in the literature. However, deep learning models are considered black boxes as they do not give an insight into which facial features of the image are considered for the prediction. Does the mask also contribute to the prediction? This study aims to unveil the behaviour of various state-of-the-art deep-learning models used for masked face recognition. For this purpose, we trained various deep-learning based masked face recognition models using the simulated masked images of CASIA-WebFace and the LFW datasets. FaceNet-based model stands out from the existing state-of-the-art with a 5-fold cross-validation accuracy of 0.93±0.005 at 95% confidence interval. Further, an explainable AI tool, SHAP, has been employed to better understand which features contribute to models’ predictions. The SHAP-marked features indicated that the trained models did not prioritize the occluded face regions.