Comparison of Face Recognition Accuracy of ArcFace, Facenet and Facenet512 Models on Deepface Framework
Andrian Firmansyah, Tien Fabrianti Kusumasari, Ekky Novriza Alam
Abstract
Face recognition is one of the biometric-based authentication methods known for its reliability. In addition, face recognition is also currently very concerning, especially with the growing use and available technology. Many frameworks can be used for the face recognition process, one of which is DeepFace. DeepFace has many models and detectors that can be used for face recognition with an accuracy above 93%. However, the accuracy obtained needs to be tested, especially when faced with a dataset of Indonesian faces. This research will discuss the accuracy comparison of the Facenet model, Facenet512, from ArcFace, available in the DeepFace framework. From the comparison results, it is obtained that Facenet512 has a high value in accuracy calculation which is 0.974 or 97.4%, compared to Facenet, which has an accuracy of 0.921 or 92.1%, and ArcFace, which has an accuracy of 0.878 or 87.8%. The benefit of this research is to test how high the accuracy of the existing model in DeepFace is if tested with the Indonesian dataset. In this test, Facenet512 is the model that has the highest accuracy when compared to ArcFace and Facenet. This research is expected to help DeepFace users determine the best model to use and provide references to DeepFace developers for future development.