Face Recognition Technology: How Machines Identify Human Faces

Facial recognition has rapidly advanced with deep learning, such that machines are able to identify individuals and infer many characteristics from facial photographs. Re- cent systems utilize convolutional neural networks (CNNs), and hybrid CNN-recurrent neural networks (RNNs) in an effort to learn automatically complex convolutional spatial features, and temporal dynamics when temporal dynamics are needed. Benchmark comparison of facial recognition systems will train on millions of faces. In addition to identity, computers can also recognize emotion, age and gender estimate, and provide real-time analytics. Deep learning pipelines, for instance, recog- nize with up to 96–99% accuracy on standard datasets. New technologies drive mass applications such as smart door locks, emotion-based music playback, and mobile device authentication. Challenges come in the form of illumination, occlusion, pose, and demographic bias. We cover recent deep face recognition and face analytics techniques in the survey, compare models and results, and provide an agenda of proximal applications (e.g. security, HCI, marketing) and open questions. We wrap up key methodologies and results. From cross-comparison, we infer trends in methodologies (e.g. CNN families, real-time libraries) and directions of future work (e.g. fairness and multimodal techniques). The scope of our review is comprehensive and seeks to provide academically equal coverage of face recognition technologies and their applications.