Face Recognition-Based Room Access Security System Prototype using A Deep Learning Algorithm

Abstract

Currently, security systems use conventional security methods, which provide low levels of security. Therefore, some organizations now use biometric-based security systems, which include facial recognition-based systems. However, processing facial data requires computationally intensive feature extraction, making real-time implementation difficult. Additionally, most data used are from public datasets. In this study, we developed a facial recognition-based security system for door access using a convolutional neural network (CNN) for real-time face recognition. We used the primary data of 102 students. The datasets include two settings (i.e., outdoor and indoor) and three facial expressions (i.e., normal, smiley, and sleepy), amounting to 3060 samples. The training was performed using three deep-learning CNN architectures: Xception (model X), VGG16 (model Y), and modified VGG16 (model Z). The best accuracy results of the three training architectures of model X, model Y, and model Z for 100 epochs are 0.9469, 0.9971, and 1, respectively. In tests conducted on the 102 test data points, models X, Y, and Z achieved accuracies of 50%, 97.05%, and 97.05%, respectively. These results indicate that the modified VGG16 (model Z) is the best for real-time testing. In real-time tests conducted on the security system prototype with 15 respondents, the resulting accuracy of model Z is 86.6%. This demonstrates that the modified VGG16 model has excellent recognition capabilities and can be implemented as a room access security system.