CMD: Real-Time Compliant Mask Detection using Transfer Learning
Document Type
Conference Proceeding
Publication Date
7-21-2023
Abstract
Wearing masks has served as one of the key practices to contain the spread of COVID-19. This study aims to offer an enhanced approach to the automated monitoring of mask-wearing compliance by developing models that identify correctly masked, incorrectly masked, occluded unmasked, and non-occluded unmasked faces through transfer learning and deploying them in real time. A curated dataset of 1200 images with equal representation of all four classes was first prepared by selecting and relabeling images from publicly available datasets such as MAFA, WIDER FACE, and MaskedFace-Net. Transfer learning was then performed on the pre-Trained models MobileNetV3 Small, ResNet50V2, VGG16, Xception, and YOLOv5 Small Classification. Upon model evaluation, YOLOv5 Small Classification emerged as the most balanced model with the second-best inference speed (23.0 ms) and a relatively high accuracy (87.78%). For the real-Time deployment, ResNet50V2 had the best overall performance, having mostly accurate detections and obtaining the second-best FPS value (4.53). Future work may involve deployment in embedded systems and exploring multi-face classification.
Recommended Citation
Pamela Anne C. Serrano, Jhorcen P. Mendoza, Ivan George L. Tarun, Vidal Wyatt M. Lopez, and Patricia Angela R. Abu. 2023. CMD: Real-Time Compliant Mask Detection using Transfer Learning. In Proceedings of the 2023 International Conference on Robotics, Control and Vision Engineering (RCVE '23). Association for Computing Machinery, New York, NY, USA, 54–58. https://doi.org/10.1145/3608143.3608153