Comparison of the Effectiveness of Deep Learning Methods for Face Mask Detection


TRAITEMENT DU SIGNAL, vol.38, no.4, pp.947-953, 2021 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 38 Issue: 4
  • Publication Date: 2021
  • Doi Number: 10.18280/ts.380404
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, PASCAL, Business Source Elite, Business Source Premier, Compendex, zbMATH
  • Page Numbers: pp.947-953
  • Keywords: CNN, deep learning, face mask detection, transfer learning
  • Gazi University Affiliated: Yes


The usage of mask is necessary for the prevention and control of COVID-19 which is a respiratory disease that passes from person to person by contact and droplets from the respiratory tract. It is an important task to identify people who do not wear face mask in the community. In this study, performance comparison of the automated deep learning based models including the ones that use transfer learning for face mask detection on images was performed. Before training deep models, faces were detected within images using multi-task cascaded convolutional network (MTCNN). Images obtained from face mask detection dataset, COVID face mask detection dataset, mask detection dataset, and with/without mask dataset were used for training and testing the models. Face areas that are detected with MTCNN were used as input for convolutional neural network (CNN), MobileNetV2, VGG16 and ResNet50. VGG16 showed best performance with 97.82% accuracy. MobileNetV2 showed the worst performance for detecting faces without mask with 72.44% accuracy. Comparison results show that VGG16 can be used effectively to detect faces without mask. This system can be used in crowded public areas to warn people without mask that may help the reduce the risk of pandemic.