Performance Enhancement of Deep Learning Based Face Detection System

Masks on Face detection and Masked face recognition applications are now crucial in many industries, given the ongoing COVID crisis where we need to wear masks all the time. Both are problematic with a traditional face recognition approach where masks create occlusion in the input affecting performance of recognizers. This paper introduced a new method for masked face recognition by Principal Component Analysis (PCA) reconstruction. With the help of PCA, we can “correctly” recover auto recognized face even if some occlusion occurs in several parts. In the method presented, PCA is first trained over a dataset of face images without reference masks to learn only about principal features and information contained throughout variations. In the recognition phase, we can project a masked face onto PCA subspace and reconstruct its low half of face with principal component. The resulting image will be compared to the database of people and their identities. Extensive experiments are conducted and show that our PCA-based reconstruction approach remarkably boosts the recognition performance in all scenarios, establishing its practicability to mask face recognition.