Face recognition is an effective method in optimizing the efficiency level of the attendance tracking system. To improve the accuracy of the student attendance system based on face recognition for master's students in mathematics at Universitas Gadjah Mada, this research uses a convolutional neural network (CNN) approach on a dataset of 747 unique student images with 15 labels that were augmented into 15,687 images. The CNN architectures tested include VGG-16, GoogLeNet, AlexNet, and ResNet-152. The results show that ResNet-152 architecture provides the best performance in face recognition with an accuracy of 99.96%, followed by GoogLeNet (99.41%) and VGG-16 (99.32%). The superiority of ResNet-152 is influenced by the ability of its residual structure in deep learning without degradation. GoogLeNet utilizes batch normalization and factorization to reduce the number of parameters and speed up training, while VGG-16 improves non-linearity with a smaller kernel. The AlexNet architecture produced the lowest accuracy of 94.10%. Although this architecture is popular, it requires more time to achieve sufficient accuracy as AlexNet has limited model depth compared to ResNet-152, GoogLeNet, and VGG-16.
Copyrights © 2026