Masked Face Recognition using Deep Learning based on Unmasked Area

Researches related to face recognition find new problems, especially due to current lifestyle change by wearing a face mask. A masked face means that features or information contained in the face are missing. Hence, some modifications should be made to the existing face recognition system in this new situation. On the other hand, some studies indicate that face recognition using partial faces still gives high accuracy, especially on the top-half face (part of the face that contains the forehead to before the nose). Therefore, the proposed training technique uses only partial faces for masked face recognition. The used dataset in this research is our own dataset consisting of 125 subjects. This dataset contains videos taken of people's faces with different devices and backgrounds. Some methodologies from previous studies were used in this study, namely YOLOv4 as a face detection method, pre-trained VGGFace model as a feature extraction method, and artificial neural network as a classification method. Using the same training technique as the previous research, our masked face recognition system achieved a test accuracy of 79.58%. When using the proposed training technique, the obtained test accuracy was 99.53%. It shows that the performance is much better when using a more appropriate training technique in the case of masked face recognition.