Face recognition, or facial recognition, is one of the largest areas of research within computer vision. We can now use face recognition to unlock our mobile phones, verify identification at security gates, and in some countries, make purchases. With the ability to make numerous processes more efficient, many companies invest into the research and development of facial recognition technology. This article will highlight some of that research and introduce five machine learning papers on face recognition.
With a multitude of real-world applications, face recognition technology is becoming more and more prominent. From smartphone unlocking to face verification payment methods, facial recognition could improve security and surveillance in many ways. However, the technology also poses several risks. Numerous face spoofing methods could be used to fool these systems. Therefore, face anti-spoofing is essential to prevent security breaches.
In order to support face anti-spoofing research, the authors of this paper introduce a multi-modal face anti-spoofing dataset named CASIASURF. As of the writing of this paper, it is the largest open dataset for face anti-spoofing. Specifically, the dataset includes 21,000 videos taken of 1,000 subjects in RGB, Depth, and IR modalities. In addition to the dataset, the authors present a novel multi-modal fusion model as a baseline for face anti-spoofing.
Published / Last Updated – April 1st, 2019
Authors and Contributors – Shifeng Zhang (NLPR, CASIA, UCAS, China) , Xiaobo Wang (JD AI Research), Ajian Liu (MUST, Macau, China), Chenxu Zhao (JD AI Research), Jun Wan (NLPR, CASIA, UCAS, China), Sergio Escalera (University of Barcelona), Hailin Shi (JD AI Research), Zezheng Wang (JD Finance), Stan Z. Li (NLPR, CASIA, UCAS, China).
In this paper, the authors present a face recognition system called FaceNet. This system uses a deep convolutional neural network which optimizes the embedding, rather than using an intermediate bottleneck layer. The authors state that the most important aspect of this method is the end-to-end learning of the system.
The team trained the convolutional neural network on a CPU cluster for 1,000 to 2,000 hours. They then evaluated their method on four datasets. Notably, FaceNet attained an accuracy of 99.63% on the famous Labeled Faces in the Wild (LFW) dataset, and 95.12% on the Youtube Faces Database.
Published / Last Updated – June 17th, 2015
Authors and Contributors – Florian Schroff, Dmitry Kalenichenko, and James Philbin, from Google Inc.
As of the writing of this article, current embedding methods used for face recognition are able to achieve high performance in controlled settings. These methods work by taking an image of a face and storing data about that face in a latent semantic space. However, when tested in fully uncontrolled settings, the current methods cannot perform as well. This is due to instances where facial features are absent from or ambiguous in the image. An example of such a case would be face recognition in surveillance videos, where the quality of the video may be low.
To help address this issue, the authors of this paper propose Probabilistic Face Embeddings (PFEs). The authors propose a method for converting existing deterministic embeddings into PFEs. Most importantly, the authors state that this method effectively improves performance in face recognition models.
Published / Last Updated – August 7th, 2019
Authors and Contributors – Yichun Shi and Anil K. Jain, from Michigan State University.
In this study, researchers from SenseTime Research, the University of California San Diego, and Nanyang Technological University studied the effects of noise in large-scale face image datasets. Many large datasets are prone to label noise, due to their scale and cost-effective nature. With this paper, the authors aim to provide knowledge around the source of label noise and the consequences it has in face recognition models. Additionally, they aim to build and release a clean face recognition dataset titled IMDb-Face.
Two of the main goals of the study were to discover the effects of noise on final performance, and determine the best strategy to annotate face identities. To do so, the team manually cleaned two popular open face image datasets, MegaFace and MS-Celeb-1M. Their experiments showed that a model trained on just 32% of their cleaned MegaFace dataset and 20% of the cleaned MS-Celeb-1M dataset achieved similar performance to models trained on the entirety of the original uncleaned datasets.
Published / Last Updated – July 31st, 2018
Authors and Contributors – Fei Wang (SenseTime), Liren Chen (University of California San Diego), Cheng Li (SenseTime), Shiyao Huang (SenseTime), Yanjie Chen (SenseTime), Chen Qian (SenseTime), and Chen Change Loy (Nanyang Technological University).
Numerous studies have been done on deep convolutional neural networks for facial recognition. In turn, numerous large-scale face image datasets have been created to train those models. However, the authors of this paper state that the previously released datasets do not contain much data on pose and age variation in faces.
In this paper, researchers from the University of Oxford introduce the VGGFace2 dataset. This dataset includes images which have a wide range of age, ethnicity, illumination, and pose variations. In total, the dataset contains 3.31 million images and 9,131 subjects.
Published / Last Updated – May 13th, 2018
Authors and Contributors – Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman, from the Visual Geometry Group at the University of Oxford.
Hopefully, the face recognition research papers above helped strengthen your understanding of the work being done in the field.
New studies in face recognition are done every year. To keep up with the latest machine learning papers and other AI news, please see the related resources below.