In this paper, we introduce a facial recognition system comprising two key components: face detection and localization, and facial recognition. For face detection and localization, the RetinaFace method is employed to accurately identify facial regions within images and to separate them from intricate backgrounds, thus facilitating facial detection based on isolated facial imagery. In the domain of facial recognition, we address the limitations of conventional convolutional neural networks (CNNs), which are typically constrained to recognizing known categories. To overcome this limitation, in our study, we leverage a Siamese network rooted in metric learning as the central architecture for facial recognition. The primary objective of this architecture is to acquire image features. It operates by minimizing the feature distance between similar images and maximizing the feature distance between dissimilar ones. Consequently, images can be directly fed into the Siamese network to extract corresponding features, followed by similarity calculation to ascertain their presence within the database. Diverging from the conventional approach of directly classifying individuals using models, we significantly inhibit the need for model retraining owing to personnel changes in the differentiation of members and nonmembers. Furthermore, the model does not increase in size with the growth of the personnel dataset. The study outcomes demonstrate that the attained average values for accuracy, recall rate, precision, and F1-Score all surpass 96%. These results robustly demonstrate the feasibility and superior performance of this approach.