© 2024 Slimane Ennajar and Walid Bouarifi. This open-access article is distributed under a Creative Commons Attribution (CC-BY) 4.0 license. Journal of Computer Science Original Research Paper Deep Transfer Learning Approach for Student Attendance System During the COVID-19 Pandemic Slimane Ennajar and Walid Bouarifi Mathematical Team and Information Processing, National School of Applied Sciences, Safi Cadi Ayyad University, Marrakech, Morocco Article history Received: 15-01-2023 Revised: 03-03-2023 Accepted: 02-09-2023 Corresponding Author: Slimane Ennajar Mathematical Team and Information Processing, National School of Applied Sciences, Safi Cadi Ayyad University, Marrakech, Morocco Email: slimane.ennajar@ced.uca.ma Abstract: Marking students' attendance has been a challenge during the COVID-19 pandemic. It is a time-consuming task due to the abnormally high number of students present during a learning session; many studies have been proposed to improve the system. However, there are still issues with each of these systems; we have employed deep transfer learning techniques using six pre- trained convolutional neural networks. We created a dataset of faces with masks and we used this dataset to assess six Convolutional Neural Network (CNN) models. We increased the training samples to improve the performance of the pre-trained models. The latter allows us to build a masked face recognition model of learners during a learning session. Due to the COVID-19 pandemic, students don facemasks to safeguard their own well-being and mitigate the spread of the virus. This has created a problem that did not exist before. The experimental findings reveal that pre-trained models, specifically caption and InceptionResNetV2, exhibit outstanding proficiency in precisely identifying faces with masks and require minimal training time. Keywords: CNN, Computer Vision, COVID-19, Deep Transfer Learning, Student Attendance System of Absence Records by Using Facial Recognition to Detect and Identify Students' Faces Introduction An intelligent system that detects students' presence in class is an application that enables teachers and school administration to automate the process of recording attendance during learning sessions and exams. This system facilitates the administration of attendance in educational institutions through the utilization of face detection and facial recognition technology. The system was developed to allow teachers to limit the time spent on registering student absences. The interaction with students during registration, in addition to saving time on manual and repetitive work, minimizes loss of productivity. During the Covid-19 pandemic, the mandatory use of masks became prevalent in areas where people gather, particularly in educational institutions and training centers. This system enables the detection of students' faces while wearing masks and performs face detection and facial recognition. The intelligent student attendance system scans documents and optimizes the management of absence records by using facial recognition to detect and identify students' faces. This study presents an approach for recording student attendance in a classroom utilizing deep transfer learning. The primary achievements of this study can be encapsulated as follows: 1) A data processing stage involves employing the MaskTheFace technique to apply a virtual mask to each image in the pins face recognition dataset. Furthermore, a data augmentation method is utilized to enhance the training samples 2) The use of six pre-trained convolutional neural networks (Xception, InceptionResNetV2, MobileNetV2, DenseNet201, ResNet101V2, and EfficientNetB0) for masked face identification 3) Assessment of the effectiveness of the six networks in identifying student attendance in a classroom learning session amid the COVID-19 pandemic, accomplished through the implementation of transfer learning The pre-trained convolutional neural networks used in this study exhibit good predictive performance. These Convolutional Neural Networks (CNNs) enable the substitution of the pre-trained network's final "fully-