Enhancing Face Mask Detection using ResNet-50 and MobileNetV2: A Transfer Learning Application

Abstract

This research is essential for organizations enforcing strict regulations for airborne diseases. This study addresses the challenge of accurately detecting individuals wearing facemasks, a task complicated by the partial occlusion of facial features during the COVID-19 pandemic. This research aims to enhance facemask detection accuracy using a two-stage transfer learning approach with pre-trained convolutional neural network models, specifically ResNet-50 and MobileNetV2. A dataset comprising 15,005 annotated images, balanced between masked and unmasked individuals, was utilized. Data augmentation techniques, structured annotation, and systematic model fine-tuning were employed to optimize performance. The system achieved notable results, with 99.1% accuracy for detecting facemasks and 98.8% for identifying individuals without facemasks. The highest recorded performance metrics under the two-stage transfer learning approach included an accuracy of 98.21%, recall of 98.52%, and specificity of 98.01% from ResNet50. In comparison, MobileNetV2 achieved the highest precision at 98.33% and a Matthew’s correlation coefficient of 98.04%. A t-test revealed a significant improvement (p < 0.00001, t = -50.88), with the two-stage transfer learning method yielding a 5.47% higher accuracy than the conventional approach. These findings demonstrate the effectiveness of advanced transfer learning techniques in improving public health monitoring systems and offer valuable insights for developing automated surveillance solutions.