Identity prediction with uncovered facial features while wearing mask

• Main Author: Koh, Ronald Lee Xiang
• Format: Final Year Project / Dissertation / Thesis
• Published: 2023
• Subjects: Q Science (General); T Technology (General)
• Online Access: http://eprints.utar.edu.my/5778/1/fyp_CS_2023_KRLX.pdf; http://eprints.utar.edu.my/5778/

Since the Covid-19 pandemic broke out in 2019, our lives had been greatly impacted and problems had arisen from different angles. People must follow a standard operating procedure to control the spread of disease. One of the noticeable changes in behaviour was most people wear a face mask to decrease the infection of the disease. However, the action of wearing a mask had disrupted the usual face recognition process. In this project, a masked face recognition system was developed to tackle the problem mentioned. The task of building a masked face recognition had been broken down into steps, which include face detection, face embedding, face classification, and face verification. Each step was dealt with individually with a specific solution. Dataset acquired for this project includes self-collected data, LFW dataset, CelebA dataset, and GMF dataset. After trial of error though experiments, the final system was developed using OpenCV HaarCascade, FaceNet, SVM, and Euclidean distance. The developed system was able to achieve a great performance of 100.00 training accuracy and 99.787 testing accuracy on known identities. While maintaining a high accuracy for known identities, the system had also achieved a low FAR of 0.0152%, 0.0006%, and 0.0038% from CelebA, LFW, and GMF dataset respectively. The time taken for the system to inference a face image was 109.8 millisecond. When implementing the masked face recognition system in webcam, it was able to recognise the known identities while the presented face was unmasked or masked. Moreover, it was also capable of robustly distinguishing known identities with unknown identities. However, the developed system was not completely perfect, it was unable to recognise multiple identities at once in one capture, does not support integration on other devices, and unable to tell whether the face presented is in its physical form or not. Overall, the system had achieved a decent performance at recognising both unmasked and masked face, but further improvements can be implemented onto the system.