A TinyML Model for Real-Time Mask Detection on Embedded Machine Learning Devices

Mask detection has emerged as an essential technology for ensuring safety and compliance in both public and private settings. This paper presents an end-to-end application utilizing Vision Wake Words (VWW) within the realm of Tiny Machine Learning (TinyML) to detect mask adherence in real-time. Capitalizing on the efficient MobileNetV2 architecture with Depthwise Separable Convolutions, our model minimizes memory foot- print and computational demands on Microcontroller Units (MCUs). We employ Transfer Learning to expedite the adaptation to mask detection tasks, significantly reducing training time and data requirements. Sensor characteristics and data preprocessing are carefully considered to manage the high data volume inherent to the ML workflow. Our original model, sized at 8.62MB, was effectively compressed to 214KB using TensorFlow Lite with Dynamic Range Quantization. This drastic reduction allows the deployment of our TinyML model on constrained devices, offering a highly scalable and cost-effective approach to enhance public health safety. Achieving a 99.6% accuracy rate in mask detection, our model underscores the precision of TinyML applications in public settings. The paper contributes to the objective of Healthcare 4.0 by enhancing healthcare services with smart sensors and IoT integration, suggest- ing that deploying TinyML for mask detection provides a promising avenue for public health safety with minimal hardware constraints.