Improved YOLO-v5 model for boosting face mask recognition accuracy on heterogeneous IoT computing platforms

Abstract

The World Health Organization (WHO) claimed that wearing masks can effectively reduce the spread of COVID-19. In this paper, we devise novel YOLO-v5s detectors for improving the accuracy of recognizing face masks under heterogeneous IoT computing platforms. First, a module of coordinate attention is merged with the original YOLO-v5s backbone to enhance the model's attention to key information and filter out redundant information. Second, the original feature pyramid network (FPN) in the neck network is replaced with the bidirectional FPN for fast multiscale feature fusion. Third, the block of adaptive spatial feature fusion is embedded into the head layer to further solve the problem of feature inconsistency with different scales. Last, the SCYLLA-IoU metric serves as the new bounding box loss function to accelerate the model convergence. The training results show that applying the improved YOLO-v5s model to the AIZOO dataset yields an improvement in recognition accuracy of +2.2% and that the improved model has lower false detections and better detection of dense crowds. In addition, we propose a new evaluation standard of the scale limit factor to measure the detection performance of the model with different shooting distances. The proposed model is deployed on three heterogeneous IoT platforms, the Google Colab cloud, a personal computing terminal, and an Nvidia Jetson Nano edge computing device, to validate the inference feasibility, effectiveness, and robustness.