A More Robust Face Mask Detection System With Combined Frontal and Angled Viewed Faces
Date of Award
12-1-2022
Document Type
Thesis
Degree Name
Master of Science in Computer Science, Straight
First Advisor
Patricia Angela R. Abu, PhD
Abstract
One protocol currently enforced by the Philippine government to combat COVID-19 is the mandatory use of face masks in public places. The problem however is that ensuring people follow this protocol is difficult to monitor during a pandemic due to other conflicting health protocols like social distancing and workforce reduction. This study therefore explores on the creation of deep learning models that consider both frontal and side view images of the face for face mask detection as a prerequisite for measuring compliance to aid in modelling COVID-19 projections. In doing so, improvements to robustness were found when compared to using models that were trained on purely frontal images. This was accomplished by first relabeling a subset of images from the FMLD dataset to be able to distinguish whether a person is wearing a Medical Mask, Non Medical Mask, or No Mask either from the front or side view. These images were then split into train, validation, and test sets. Four deep learning models (YOLOv5 Small, YOLOv5 Medium, CenterNet Resnet50 V1 FPN 512x512, CenterNet HourGlass104 512x512) were later trained on the training set of images. These four models were compared with three models (MobileNetV1, ResNet50, VGG16) that were trained on purely frontal images. Comparisons were performed with respect to classification accuracy on the combined validation and test sets. Results show that the four models trained on the relabeled FMLD dataset offer an approximately 20% increase in accuracy over the three models that were purely trained on frontal images. Further insights include the models being able to detect multiple faces, but the accuracy decreases as more faces are present. Improvements to the dataset can be further carried out as well to include more skin complexions since the relabeled FMLD dataset is biased towards light-skinned people.
Recommended Citation
Tarun, Ivan George L., (2022). A More Robust Face Mask Detection System With Combined Frontal and Angled Viewed Faces. Archīum.ATENEO.
https://archium.ateneo.edu/theses-dissertations/980
