iFR: A New Framework for Real-Time Face Recognition with Deep Learning

Abstract

This paper investigates the performance of the real-time face recognition system with machine learning, as well as the performance of each haarcascade classifiers that based on accuracy and speed. We employ the subset of machine learning called deep learning to the real-time face recognition system as the deep face recognition technique has improved the state-of-the-art performance. This system uses a pre-trained model named FaceNet and employs triplet loss technique to impose a margin between every pair of faces from the same person to other faces. In other words, it minimizes the distance between the anchor and the positive from the same identity and maximizes the distance between the anchor and the negative from different identity. Furthermore, to further investigate the performance of the system, we implement the Tensorflow framework to improve the system performance by the usage of Graphics Processing Unit (GPU). This system uses Labeled Faces in Wild (LFW) dataset as the benchmark to test the performance of the face recognition system. Apart from that, we conduct a preliminary experiment to evaluate the performance of haarcascade classifiers so that we can choose the best classifier in term of accuracy and speed. Haarcascade_frontalface_default.xml (FD) exhibits best performance compared to haarcascade_frontalface_alt.xml (FA) and haarcas-cade_frontalface_alt2.xml (FA2) with accurate number of faces detected and shortest average time taken to detect faces.