Masked and Unmasked Facial Recognition: Technical Appraisal of a Conventional Prototype and a Proposed Deep Learning Architecture
K. W. I. P. Dissanayaka *
Department of Statistics & Computer Science, University of Kelaniya, Kelaniya, Sri Lanka.
*Author to whom correspondence should be addressed.
Abstract
Aim: To appraise the documented performance and limitations of a masked/unmasked facial-recognition prototype and to specify a testable deep-learning architecture for subsequent evaluation.
Design: A technical appraisal of thesis-reported prototype outcomes and a proposed but unimplemented architecture; no new comparative experiment is reported.
Methods: The historical system combined an OpenCV deep-learning face detector with support vector machine (SVM)-based identity classification. The proposed system combines landmark-assisted face detection and alignment, ArcFace-style embeddings, cosine-similarity matching, an unknown-identity decision and optional ONNX Runtime inference. Evaluation requirements are defined for identity partitioning, mask conditions, multiple-person scenes and error reporting.
Results: The thesis account attributes a face-detection accuracy value of 99% and approximately 80% live identity-classification accuracy to the prototype, while also presenting values of 97% for masked faces and 87% for unmasked faces elsewhere. Underlying test counts and predictions are unavailable, and these figures cannot be reconciled or recalculated. No accuracy or latency findings are available for the proposed architecture.
Conclusion: A detector–embedding–matching workflow offers a technically justified route for further study, but its effectiveness, fairness, operating speed and suitability for deployment remain unestablished without reproducible evaluation.
Keywords: Facial recognition, face masks, face detection, ArcFace, RetinaFace, face embeddings, support vector machine, biometric evaluation