Facial Detection with mask

Abstract

Since the advent of COVID-19 pandemic the problem of masked facial recognition has seen tremendous spike in both practice and academia. Existing state-of-the-art facial recognition models have struggled to accurately classify obscured faces, highlighting the pressing need for a comprehensive and resilient facial recognition framework. This demand stems from various practical applications, including surveillance, identity ver ification, ticket generation for parks, and attendance systems, necessitating a model that is both robust and efficient for real-time operation. This thesis addresses the need for such a solution by developing and validating a robust, time-efficient system tailored for real-time applications on Embedded AI hardware. In today’s world, where mask-wearing has become ubiquitous for public health reasons, ensuring accurate and efficient facial recognition in such scenarios is paramount. Moreover, the practical ap plications of facial recognition extend beyond security and authentication to include personalized services and access control in various domains such as retail, banking, and healthcare. Through comprehensive analysis, various architectural approaches have been thor oughly examined to address the challenges of masked facial recognition. An emphasis has been placed on employing a multi-headed attention mechanism, aimed at prioritiz ing unobscured facial regions. The choice of loss functions is to optimize the geodesic distance among embeddings of the masked faces. Careful consideration has been given to computational efficiency, ensuring the algorithm’s suitability for real-time applica tions. Additionally, a diverse range of model scopes has been explored, including the utilization of minimal image features and employing generative models to unveil ob scured faces. Following training, various optimization techniques are applied to enhance the system’s performance on Edge AI devices, particularly leveraging the capabilities of platforms like the Jetson Xavier AGX.