Article Preview
TopIntroduction
Facial recognition systems like other biometric systems have become ubiquitous and have touched various aspects of the human-computer interaction in this digital age. Facial recognition systems (FRS), when used in collaboration with advanced computing including artificial intelligence, provide a multitude of opportunities which encompass security, feature detection, emotional status identification, law enforcement, personal payment systems amongst others (Shein, 2011; Truong, Diep, & Zelinka, 2020). Additionally, the technology convergence in the smartphone has led to the development of “data-hungry” apps that have cut across national borders (O’Flaherty, 2019) and thus more avenues for data collections for FRS have opened up. These trends are compounded by the lack of awareness, security misconceptions and concerns of the end-user (Piccolotto & Maller, 2014; Zimmermann & Gerber, 2017). FRS have long relied upon building facial image databases that are proprietary in nature that took extensive resources to build and maintain. However, with the advent of social media and other publicly accessible image repositories, the resource level challenges in building the FRS have been significantly reduced. Additionally; data collection, storage and analysis intermediaries have emerged who can provide “FRS-as-a-service” thereby bringing down the barriers for adoption as well. However, the recent data breach of a prominent US firm (Clearview AI) has thrown light upon the systematic risks of “FRS-as-a-service” (O’Flaherty, 2020; Valinsky, 2019).
FRS is a specialized version of image recognition systems that can be built either to identify the face as a whole or extract features from the image thereby identifying the face using the extracted features (Abbass & Ibrahim, 2013; Bansal, Agarwal, Sharma, & Gupta, 2013). The facial image in 3D or 2D is captured using a camera and is then compared with the face database to judge whether the face is recognised or not. The generalised steps involved in facial recognition are presented in Figure 1.
Figure 1. Generalised Steps in Facial Recognition Systems
There have been significant researches conducted to improve the effectiveness and efficiency of facial recognition at every step identified in Figure 1. The advances in detection of the image include feature level detection like eye detection (Bianchini & Sarti, 2006) and innovative functions like liveness and emotion detection (Nishanth & Rao, 2019; Reney & Tripathi, 2015; T. Wang, Yang, Lei, Liao, & Li, 2013; Yin et al., 2017). Image processing advancements include functions like correcting face alignment (L. Zhang, Allebach, Lin, & Wang, 2015), head position variations (Biswas, Aggarwal, Flynn, & Bowyer, 2013; Shiau, Pu, & Leu, 2010; Shinwari, Kosunen, Alariki, & Naji, 2019; C. Wang & Brandstein, 2000; L. Zhang et al., 2015) and lighting fluctuations (Doi, Sato, & Chihara, 1998).
The improvements in the “matching” step include the use of machine learning and other algorithmic developments to improve the performance even in compromised environmental scenarios (Biswas, Bowyer, & Flynn, 2010; Hirano, Ikeda, & Nakamura, 2003; Li, Lu, & Li, 2017; Megreya, 2018; X. Wang & Guan, 2018). Similar, algorithmic improvements have been identified in the “verification step” as well (Alfarsi, Jabbar, Tawafak, Alsidiri, & Alsinani, 2019; Payne, Solheim, & Castain, 1993; Peng & Shen, 2013; Ramanathan & Chellappa, 2005; Taheri, Mozaffari, & Keshavarzi, 2015).