Linear Transformations in Masked Face Recognition: Bridging the Representation Gap

Amid the COVID-19 pandemic, the ubiquity of masks revealed the limitations of face recognition systems in accurately identifying individuals with occluded faces. This issue goes beyond pandemics and applies to circumstances where face coverings are compulsory, such as in healthcare establishments or industrial locations. Optimizing face recognition performance for masked faces is imperative for ensuring security and safety in diverse environments. Earlier endeavors to overcome this challenge faced hurdles such as limited datasets of masked faces, which made it arduous to adjust existing face recognition models to the exclusive traits of masked faces. This study introduces a cutting-edge empirical linear transformation technique for face recognition, modifying embedded vector representations of unmasked faces to correspond with masked faces and vice versa. The study's outcomes show a firm linear relationship between masked and unmasked face embeddings, emphasizing the potential for knowledge exchange between these two facial conditions. Furthermore, this innovative linear transformation strategy offers an efficient and effective solution for masked face recognition systems, eliminating the need for retraining on extensive masked face datasets.