An Incremental Training on Deep Learning Face Recognition for M-Learning Online Exam Proctoring

Abstract

The ability to provide an academic resource for a remote student has increased the use of m-learning in distance education. Online exams as a tool to measure the student's outcome need a proctoring method to detect cheats. Several methods had been proposed to fulfill these needs, from a no-proctoring exam to automatic online supervision. A visual verification during an online exam is required to verify a student took the exam, therefore a CNN-FR is used to do it. The problem that exists in face recognition is the system invariant against pose and lighting variations. In some proposed methods, an additional process such as image equalization and SURF features is executed to overcome the problems. In this paper, we proposed an incremental training process on face recognition training, so there will be no need to add another process so it will reduce the computation cost and time. To acquired high accuracy we've analyzed four different face detectors, which are Haar-cascade, LBP, MTCNN, and Yolo-face, as in face recognition a Facenet model was tested. The evaluation of the proposed method shows that a deep learning face detector has overcome the others, on the other hand, an incremental training of facenet model results in a smaller dataset size by 1% with a faster training time of 7% on Yolo-face face detector and 64% on MTCNN compared to batch training. The proposed method results in an equally high accuracy rate as in batch training (98%).