Good Generalization on Face-Mask Detection Based on Simple Convolutional Neural Network

ABSTRACT

Coronavirus pandemics have influenced people's daily life seriously since 2019. Authorized organizations suggested people wear a mask in public areas can significantly reduce the probability of getting infected. Thus, we proposed a method based on a simple convolutional neural network (CNN) to perform mask detection. The whole developing process was divided into two stages and mainly used three datasets (dataset_l, dataset_2 and dataset_3). Dataset_1 has images of people with and without masks. Dataset_2 and dataset_3 have one more category-images of people wearing masks incorrectly. The first stage was to train the model based on dataset_1 and it achieved 100% accuracy on validation set. It could also be applied to another two similar datasets without any training on them with accuracy 73.55% and 66.80% respectively. In the second stage, to detect people wearing masks incorrectly, the same model was trained based on dataset_2. The accuracy of this model reached 99.34%. However, when applying it directly to dataset_3, only 44.50% accuracy was achieved. To improve the accuracy, the distribution of dataset_2 and dataset_3 was rearranged. Finally, the accuracy of the model for dataset_3 was nearly 80%. We concluded that generally deep learning models would have better generalization on mask-detection tasks and our model was good at handling two-label mask dataset. © 2022 IEEE.