Siamese Network's Performance for Face Recognition

Abstract

Performance of Siamese network for real-time face recognition software in a one-shot learning setting is discussed in the paper. Two loss functions for the Siamese network are also compared, which are the contrastive loss and the triplet loss. Initially, a multitask cascaded neural network detects faces from a webcam, and the Siamese network matches the detected faces to the user's registered face. In the experiment evaluation, we find that the Siamese network with contrastive loss achieves better performance. The accuracy is 0.8875. However, the model with triplet loss has an accuracy of 0.85.