Siamese Model-Based Face Verification Using CNN and MobileNetV2

Face verification plays an important role in computer vision, especially in mobile and embedded systems with limited computational capacity. This study proposes a face verification system based on the Siamese Neural Network (SNN) architecture by integrating six embedding models. These models consist of a standard CNN, an L2-normalized CNN, a baseline MobileNetV2, a structurally adjusted MobileNetV2, a pre-trained MobileNetV2, and a fine-tuned MobileNetV2. The dataset includes facial images captured from three webcams and additional samples obtained from the Labeled Faces in the Wild and ImageNet datasets. The experimental procedure includes image preprocessing, construction of balanced positive and negative image pairs, model training, and evaluation using accuracy, precision, recall, F1-score, and AUC. The results show that the pre-trained MobileNetV2 and the standard CNN achieve the highest verification accuracy, reaching 100 percent and 99.998 percent, respectively. Among all models, the structurally adjusted MobileNetV2 presents the best trade-off by combining high accuracy, computational efficiency, and training stability while successfully avoiding overfitting. The real-time implementation involves only the structurally adjusted MobileNetV2 model due to its lightweight structure and consistent performance. This model produces low embedding distances, low latency, and high throughput during CPU-based inference. The performance outperforms GPU execution in one-by-one image processing. The proposed system offers a practical and efficient face verification solution for deployment in identity authentication applications on resource-constrained platforms. These findings support the development of scalable and adaptive biometric security systems that rely on deep learning.