The COVID-19 pandemic demonstrated the importance of speed and efficiency in developing algorithmic solutions for novel problems. We compare four different approaches for building classifiers for proper mask usage using a limited training dataset of 240 images. We find that though SVM and CNN models work adequately, transfer learning with pretrained AlexNet eclipses other models in performance metrics. Additionally, a Symbolic Reasoning Model using predefined facial feature detectors present novel insights on building data-efficient models. While the performance of the Symbolic Reasoning Model is only on par with SVM and vanilla CNN, the method shows potential to develop into an explainable and flexible alternative for deep learning models.
kimyoonduk/proper-mask-classification
Building an efficient and explainable model for identifying proper face mask usage (CIS 522 - Deep Learning, 2021)
Jupyter NotebookMIT