Toward a Transfer Learning Approach to Detect Face Mask Type in Real-time

Face masks are considered protective equipment that has the ability to safeguard humans from vulnerable situations. Although there exists a wide range of masks specifically designed for diverse purposes, there is a terrible lack of concern regarding proper usage. Consequently, the generalization of their usage can cause many life-threatening problems. As a result, a system that can detect the type of face mask can play a life-saving role to ensure the proper usage of these safety gear. With this aim, a custom dataset was built by manually labeling face mask images which include 8 classes. Scratch CNN and four transfer learning models have been implemented and their performance was thoroughly evaluated and assessed on multiple criteria to select the best one. Based on the investigation, it is found that SSD MobNet V2 achieved the highest accuracy of 83%. The developed system takes real-time video stream input from the camera and can detect the type of mask in different conditions.