Akhmad Ramadani, Rosa Andrie Asmara, Ulla Delfana Rosiani
This study investigates the performance of facial recognition systems by integrating various deep learning architectures and efficient nearest neighbor algorithms. We compare pre-trained CNN models such as FaceNet512 and ArcFace, combined with face detectors like OpenCV and MTCNN, to evaluate recognition accuracy and speed. Two nearest neighbor search algorithms, Annoy and Voyager, are assessed based on cosine distance accuracy and computation time. Results show that Voyager yields more accurate cosine similarity values with an average error margin near zero, while Annoy executes searches faster, averaging 0.07 seconds per query compared to Voyager's 0.08 seconds. Threshold tuning experiments reveal that a cosine similarity threshold of 0.5 provides optimal accuracy, correctly identifying 103 out of 107 test cases with only 4 misclassifications. Furthermore, a comparison of system configurations shows that the High Accuracy, Low Speed model (MTCNN + FaceNet512) achieves 100% detection and recognition, whereas the Low Accuracy, High Speed setup (OpenCV + FaceNet512) achieves only 66 successful identifications. These results highlight practical trade-offs between recognition accuracy and speed, providing guidance for deploying robust, real-time facial recognition systems in digital security applications. Using this result, one can choose which model based on user face image registered amount. © 2025 IEEE.
Politeknik Negeri Malang, Information Technology Department, Malang, Indonesia