Analysis of selected machine learning algorithms for development of face recognition system for access control in the post COVID-19 era

The mandatory wearing of face masks orchestrated by the COVID-19 pandemic has brought some complexities to the ability of face recognition systems in identifying such faces. Thus, this study examined the performance of some selected machine learning algorithms: Convolutional Neural Network (CNN), Linear discriminant analysis (LDA), Logistic Regression (LR), K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Naïve Bayes (NB) and Decision Tree (DT) for the recognition of masked and unmasked human faces; and developed a face recognition system based on the algorithm with the best performance. The dataset was composed of 11,792 masked and unmasked facial images collected from Kaggle online repository. The set was divided into 70%, 20% and 10% respectively for training, testing and validation each for the masked and the unmasked face images. The images were augmented randomly for training, with some rotated 30 degree, some zoomed 20%, some shifted horizontally by 10% in width, some shifted vertically by 10% in height and some flipped horizontally. The performance analysis of the algorithms presented accuracies of 99% and 99% for the CNN; 93% and 98% for LDA; 93% and 93% for LR; 86% and 78% for the NB; 70% and 70% for the KNN; 64% and 45% for the DT; and 92% and 98% for the SVM respectively for training and cross validation. Overall, the CNN recorded the highest recognition accuracy. Thus, a CNN user-friendly face recognition, system was developed, and tested with a number of real-life human masked and unmasked faces which showed excellent recognition performance. The implementation was carried out with appropriate Python machine libraries. The developed system could be very useful for access control and surveillance. Therefore, this study recommends the adoption of the developed systems for face recognition-based security system.