A novel facial recognition technique with focusing on masked faces

The recognition of the same faces masked and unmasked is a paramount function in preserving consistent recognition in public security, safety, and access control. Facial recognition technologies have been seriously tested with the widespread use of masks due to infectious diseases in recent years, which cover key facial areas and reduce identification levels. In this paper, we introduce a novel Masked-Unmasked Face Matching Model (MUFM) that uniquely leverages cosine similarity to match masked and unmasked face images a task that, to our knowledge, has not been addressed before. Our approach uses transfer learning with pre-trained VGG-16 for discriminative facial feature extraction followed by feature structuring using a K-Nearest Neighbors (K-NN) classifier. The most significant innovation is the utilization of cosine similarity to compare feature embeddings, such that strong identification is possible even when critical facial regions are obscured. To establish the model proposed, we have developed a comprehensive dataset from three different sources i.e., real-world pictures resulting in 95% recognition. This work not only addresses a vital gap in occluded face recognition but also offers a scalable solution to security and surveillance activities across environments with varying occlusion rates.