Publication: Deep gender and face identity recognition for attendance recording system
Loading...
Date
2020-07-01
Authors
Siong, King Soon
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Deep gender and face identity recognition attendance system helps to automate the attendance recording process in a class. The biometric technique consists of two main parts which are gender classification and face identity recognition. The gender classification is run before the recognition to reduce the misidentification made by the system. The performance recognition in terms of mean probability has shown improvement of at least 7.67%. Both gender classification and face identity recognition are performed through a deep learning approach followed by two different Support Vector Machine (SVM) classifiers with linear kernel trained. The SVM used for gender classification classified face into two classes, male and female. The output of gender classification decides whether male or female face identity recognition to be carried out in which another SVM classifier classifies the face based on the name of the student is used. The datasets used to train SVM classifiers are obtained from the students in the class. They are then prepared through face detection and alignment using Multi-task Cascaded Convolutional Network (MTCNN). The detected and aligned faces undergo augmentation that includes random crop, flip, and prewhiten before feeding to train the SVM classifier. The pre-trained model used for face feature extraction is the Inception ResNet-v1 architecture trained using the VGGFace2 dataset. Only two students’ faces which consist of a male and a female are selected to evaluate the trained SVM due to the Coronavirus disease 2019 (COVID-19) outbreak. The test data are obtained in the hostel room during the movement control order (MCO) of Malaysia using the C922 Logitech Webcam. The attendance sign-in process is executed by activating webcam to start video streaming. After face detection is done, the RGB face image is pre-processed with prewhiten before feeding to the convolutional neural network for feature extraction.