On the Performance of Face Recognition Models for Local Faces Dataset

The number of facial recognition models that have been developed makes many users free to choose the desired model. However, the large number of face recognition models will also raise questions about how effectively the model is used, especially when it is used to verify certain types of faces, for example, Indonesian faces. In this paper, the evaluation results of several face recognition models using Indonesian faces are presented. The models evaluated were facenet based on inception resnet v2 from the deepface library, VGGFace from the deepface library, and facenet Google based on inception resnet v1 from the keras-facenet library. The face detection model used for cropping images is SSD. The method used to carry out the evaluation is to compare the embedding results of two Indonesian sample faces and evaluate the actual value and the predicted result value of the calculated distance. This value is then used to calculate the confusion matrix and the accuracy, precision, recall, and f1score parameters will be calculated [1]. The results show that the keras – facenet model gets the best value with an accuracy of 96.51% with a precision of 100% in the dataset using euclidean threshold 0.75. The calculated evaluation results can be used for considerations in using the face recognition models that have been mentioned, especially for Indonesian faces.