Abstract:
In recent years, deep learning-based algorithms have been immensely employed and tested in a variety of real-world
applications. The efficacy of such algorithms has been thoroughly examined in a practical setting. In this paper, CNN-based deep
learning approaches are utilized to recognize faces in real-time to identify faces with and without mask. We employ pre-trained
algorithms (YOLOv2 and SSD) to identify people wearing a face mask, which enables a machine to perform recognition tasks while
evolving through a learning method. Meanwhile, if there is more than one person in the scene, the one with the max score will be
selected for classification. Thus, a hybrid approach that combines YOLOv2 and SSD algorithms to work in parallel is developed for
masked-face extraction. Likewise, the Viola-Jones algorithm is used here to detect faces without mask and randomly select a single
region of interest (ROI) to be stored for classification. All pre-processing algorithms work separately in parallel as reconstruction
steps for preprocessing to crop the ROI and store images for training and testing dataset. Followed by developing a lightweight
computational complexity CNN model for face mask recognition to identify whether the selected person’s face is wearing a mask or
not. The dataset contains numerous variations in appearance and viewpoint to capture different scenarios with and without mask
faces. On average, the proposed face mask detection architecture realizes recall and F1 score of 98.3 and 98.31, respectively. The
training performance, on the other hand, has improved by 19.7% and 95.9% for training time and storage space (model size)
compared to AlexNet. The presented framework architecture is an efficient face mask and unmask detector and can be employed as a
robust medical assistant face detector in the healthcare sector for automated tracking of a patient, visitor, or staff member wearing a
mask or not.
