Pynq-YOLO-Net: An Embedded Quantized Convolutional Neural Network for Face Mask Detection in COVID-19 Pandemic Era

Abstract

The recent Coronavirus COVID-19 is a very infectious disease that is transmitted through droplets generated when an infected person coughs, sneezes, or exhales. So, people must wear a face mask to reduce the power of the transition of this virus. Governments around the world have imposed the use of face masks in public spaces and supermarkets. In this paper, we propose to build a face mask detection system based on a lightweight Convolutional Neural Network (CNN) and the YOLO object detection framework to implement it on an embedded low power device. The object detection framework was designed using a single Convolutional Neural Network for object detection in real-time. To make the YOLO framework suitable for embedded implementation, we propose to build a lightweight Convolutional Neural Network and quantize it by using a single bit for weight and 2 bits for activations. The proposed network called Pynq-YOLO-Net was implemented on the Pynq Z1 platform. The computation was divided between the software and the hardware. The features extraction part was executed on the hardware device and the output part was executed on the software. This configuration has allowed reaching real-time processing with a very good detection accuracy of 97% when tested on the combination of collected datasets.