Real-Time Face Mask Detection Using YOLOv8n with Performance Optimization

Abstract

The COVID-19 pandemic has highlighted the practical limitations of manual monitoring in high-density population and emphasized the importance of complying with face mask rules in public places. The aim of this study is to present a full design, implementation, and systematic evaluation of a binary face mask identification system based on the YOLOv8n deep learning model. The dataset used for training was taken from Roboflow and labeled with two classes and 929 annotated instances. It includes 120 training shots and 128 validation images. The model was trained using the Ultralytics YOLOv8 framework on a cloud-based GPU environment for five epochs. The model was trained using this dataset. The trained model has attained a precision of 0.681, recall of 0.582, mean Average Precision at IoU threshold 0.5 (mAP50) of 0.655 and mAP50-95 of 0.490. Per-class analysis showed that the mask class achieved a mAP50 of 0.777, while the no-mask class achieved a mAP50 of 0.534. This asymmetry is attributed to the class imbalance in the training data. As per confusion matrix evaluation, 434 correct mask classifications and 216 appropriate no mask classifications are made out of total 929 validation cases. It was proven that near real-time inference can be done on standard GPU hardware at 154 milliseconds per image (~ 6.3 frames per second). In this work we provide a fully repeatable baseline, with the whole training pipeline running in 0.241 hours and using just publicly available technology. The findings are competitive to those of published systems including FMD-YOLO (66.4% VOC mAP) but with far lower implementation complexity. This shows that YOLOv8n, in conjunction with Roboflow and Ultralytics, provides an accessible and efficient pipeline for automated public health compliance monitoring.