Face mask detection and classification via deep transfer learning

Wearing a mask is an important way of preventing COVID-19 transmission and infection. German researchers found that wearing masks can effectively reduce the infection rate of COVID-19 by 40%. However, the detection of face mask-wearing in the real world is affected by factors such as light, occlusion, and multi-object. The detection effect is poor, and the wearing of cotton masks, sponge masks, scarves and other items greatly reduces the personal protection effect. Therefore, this paper proposes a new algorithm for mask detection and classification that fuses transfer learning and deep learning. Firstly, this paper proposes a new algorithm for face mask detection that integrates transfer learning and Efficient-Yolov3, using EfficientNet as the backbone feature extraction network, and choosing CIoU as the loss function to reduce the number of network parameters and improve the accuracy of mask detection. Secondly, this paper divides the mask into two categories of qualified masks (N95 masks, disposable medical masks) and unqualified masks (cotton masks, sponge masks, scarves, etc.), creates a mask classification data set, and proposes a new mask classification algorithm that the combines transfer learning and MobileNet, enhances the generalization of the model and solves the problem of small data size and easy overfitting. Experiments on the public face mask detection data set show that the proposed algorithm has a better performance than existing algorithms. In addition, experiments are performed on the created mask classification data set. The mask classification accuracy of the proposed algorithm is 97.84%, which is better than other algorithms.

Automated summary

Wearing masks is an important way to prevent the transmission of COVID-19, but detecting mask-wearing in the real world is challenging due to various factors. This study proposes a new algorithm for mask detection and classification that combines transfer learning and deep learning, showing superior performance in experiments compared to existing algorithms.