Development of a face mask detection and masked facial recognition model based on a hybrid convolutional neural network

Abstract

In the context of enhancing security authentication, facial recognition technology has become pivotal, replacing conventional authentication methods such as passwords, security tokens and PINs across various sectors. However, the rapid growth of facial recognition technology faced hindrances due to the COVID-19 pandemic, where mandatory face mask usage obscured facial features, challenging existing authentication methods. Regardless, the existence of several methods for face mask detection and recognition highlighted prevalent issues such as poor lighting, varied angles, failed detection for improper use of face masks, computational complexity, difficulty in detecting smaller faces and low-resolution targets have led to suboptimal accuracy rates. Hence, this study aims to address these challenges by introducing a hybrid Convolutional Neural Network (CNN) architecture tailored for Face Mask Detection (FMD) and Masked Facial Recognition (MFR). The models used MobileNetV2 and FaceNet InceptionResNetV1 respectively for FMD and MFR. The proposed models leverage advanced FMD and MFR technologies, contributing to the real-world need for enhanced security in scenarios where traditional methods are insufficient. The models underwent training using five distinct datasets, comprising a total of 10,980 images for FMD across two datasets, and 26,523 images for MFR across three datasets. In the FMD phase, the model achieved exceptional results, attaining a perfect 100% across evaluation metrics such as accuracy, precision, recall, and the f1-score within a training timeframe of an hour. Transitioning to the MFR phase, where the model required approximately one hour and 30 minutes, maintained an outcome of 99.68% across the aforementioned metrics, surpassing the accuracy level of existing models within the meta-analysis. Furthermore, the model underwent testing on a real-time custom dataset designed for MFR evaluation, consisting of 5500 images (i.e., 4400 for training, 550 for validation and 550 for testing) in real-life scenarios. Robustness was assessed under various conditions, resulting in an impressive 99.82% accuracy. The model demonstrated high accuracy in real-time testing. Notably, both the models excel in detecting and recognising masked participants from diverse angles and lighting conditions with minimal computational complexity. Leveraging the pre-trained MobileNetV2 for FMD and FaceNet InceptionResNetV1 with CNN for MFR, the CNN models provide a comprehensive solution. The proposed models surpass existing methods, excelling in accuracy under challenging conditions. This study contributes a versatile and efficient solution, addressing limitations in current approaches and providing robust models for FMD and MFR in diverse sectors.