Efficient Deep Learning Architecture for Facemask Detection

Since COVID-19 pandemic all over the world, wearing mask become mandatory in public space. In order to enforce the new normal behaviour, regulators need to ensure every person is wearing a mask in order to avoid the spreading of the viruses. Before the pandemic, there were a number of closed-circuit televisions (CCTV) installed in public space for security purposes. The research aims to identify algorithms with acceptable classification quality and at the same time low computing complexity. This research aims to identify the algorithm to identify Face Mask. This research uses two public datasets, the first dataset has two labels with and without mask, and the second dataset consists of three labels (facemask, improper use of facemask and proper use of facemask). This research examines some well known deep learning architectures which are VGG, MobileNet, MobileNetV2, EfficientNet B0, NasNetMobile. A modification of VGG to reduce the number of parameters is also examined. An evaluation of the classification performance and execution time in the testing set is carried out on binary and three class dataset. According to the experiments, Modified VGG with 7 layers with 1.6 Million parameters consistently achieves fastest performance. The classification performance for three class dataset is achieved by Modified VGG (CVGG-7) at 100% while for the binary facemask classification is achieved by MobileNetV2 at 99.7%