Face recognition and verification systems play a crucial role in many critical areas such as biometric security, access control, and user authentication. This study presents a training-free (zero-shot) face verification protocol and comprehensively compares the performance of different pre-trained deep learning models-Facenet-IRv1, ArcFace, ResNet-18, VGG16, AlexNet, and OpenFace-on the Labeled Faces in the Wild (LFW) dataset. In the proposed approach, two input images are passed through the same network using a Siamese-like inference process, and the resulting embeddings are compared using cosine similarity after L2-normalization. To classify the similarity scores obtained from the model outputs, dynamic threshold calibration is applied for each model, maximizing Youden's J statistic, and this threshold value (𝜏) is transferred to the test dataset without any additional optimization. Additionally, multiple metrics such as ROC-AUC curve, accuracy, precision, recall, F1-score, average inference time, and FPS were calculated to evaluate model performance independently of the threshold. The findings indicate that ArcFace and Facenet-IRv1 models surpassed others in terms of accuracy and reliability, while lightweight architectures such as ResNet-18 and VGG16 offer speed advantages, making them suitable alternatives for real-time applications. These results demonstrate that approaches that do not require training from scratch offer both a cost- and time-efficient solution in face verification systems. In this respect, the study introduces a standardized framework that enables a multidimensional evaluation of different architectures without the need for additional training and offers quantitative insights into the accuracy–speed trade-off in the field of face verification.
Mehmet Özdem (Wed,) studied this question.