FTUI Doctorate Develops Feature Extraction in Masked Face Recognition Systems

The facial recognition system that uses a classical approach so far has not been able to provide optimal results when faced with occlusion challenges. One of them is when the face is wearing a mask. If someone uses a facial recognition system application by having to open the mask first in a public place, of course, it is very dangerous for the safety and health of all parties. So that a facial recognition system is needed that has high system performance with the challenge of mask occlusion.

Regina Lionnie raised this problem in her dissertation research entitled “Development of the Curvature Method for Feature Extraction in Masked Face Recognition Systems” which was presented at the FTUI Doctoral Promotion Session (15/11).

This study aims to obtain a combination of new methods or improve existing methods and be able to identify individuals when wearing masks without reducing the performance of the recognition system accuracy. The purpose of identification when using a mask is to help people use applications and software with facial recognition system algorithms accurately without the need to open a face mask. In this study, Regina built a face recognition system with masks using a holistic and partial face approach.

“In a holistic-based approach, the recognition system will use image data of the entire face when wearing a mask (internal), as well as hair and ears (external). In the partial face approach, the recognition system uses areas not covered by masks, and key features that are robust to variation and how to represent these features will be developed. The feature extraction method used is a combination of curvature methods that use first and second-order partial derivatives with analytical methods, namely gray level co-occurrence matrix (GLCM) and multi-resolution analysts (MRA), including discrete wavelet transform (DWT), scale-space (SS) and wavelet packet transform (WPT),” said Regina.

In this study, Regina also found a new criterion (research novelty) called the curvature best basis (CBB) for selecting the basis for the best basis algorithm in WPT. The new criterion for selecting the best basis is dynamic and uses the highest value of the statistical measure of the standard deviation of the mean curvature of the wavelet coefficient. The best base works as extracted features that work inside the recognition system. Then this study was evaluated using the RFFMDS v1.0, RFFMDS v2.0 EYB, and UBIPr datasets.

“Results of Doctor Regina’s research show that a facial recognition system with a mask occlusion challenge was successfully built using a holistic approach with a system recognition accuracy of 98.11% and with a partial face approach with an accuracy of 98.80%. Both of these best accuracy results were obtained by the best basis curvature method. The performance of the recognition system using the best basis curvature method with a holistic or partial face approach shows the highest performance compared to the performance of previous studies. This is certainly an innovation that is very helpful in the screening process in public places that need facial recognition accuracy without compromising safety, health, and public convenience,” said the Dean of FTUI, Prof. Dr. Heri Hermansyah, ST., M.Eng., IPU regarding Regina’s research.

A research dissertation on feature extraction in a masked face recognition system has succeeded in bringing Regina Lionnie to a doctorate and is listed as the 147th graduate of the Electrical Engineering Department and the 475th graduate of the Faculty of Engineering, Universitas Indonesia. The Promotion Session was chaired by Prof. Dr. Ir. Harry Sudibyo S., DEA, with promoter Prof. Dr. Ir. Dadang Gunawan, M.Eng. and co-promoter Dr. Catur Apriono, S.T., M.T., Ph.D. While the testing team consisted of Dr. Basari, S.T., M.Eng.; Dr. Eng Mia Rizkinia ST., MT.; Siti Fauziyah Rahman, S.T., M.Eng. Ph.D.; and Dr. Ian Yoseph, ST., MT.


Bureau of Public Communications
Faculty of Engineering, Universitas Indonesia