How to strike a good balance between the merits and demerits of facial recognition becomes an important topic for vendors and users alike.
Facial recognition is seeing increasing demand and
acceptance due to its various benefits. Yet at the same time, as we pointed out throughout last month, it can trigger worries and concerns as well. Indeed this is the kind of love-hate relationship that we find ourselves in with facial recognition. How to strike a good balance between its merits and demerits, then, becomes an important topic for vendors and users alike.
Facial recognition has certain merits. One of its biggest advantages is accuracy. Over the years, facial recognition has become more accurate than ever. “The accuracy of facial recognition algorithms has increased exponentially over the years. There are now many companies which are able to claim they have a 98-99 percent false-non-match-rate if the system is used under perfect conditions and with a powerful server,” said Charlie Bennett, VP SAFR Europe.
Two factors have made this possible: one is advances in
deep learning and AI, and the other is the availability of more powerful GPUs.
“Deep learning applies multiple processing layers to learn representations of data with multiple levels of feature extraction. Facial recognition can leverage the hierarchical architecture in deep learning algorithms to learn discriminative face representation and have dramatically improved state-of-the-art performance for real-world applications. With the availability of powerful GPUs and deep learning algorithms, and large sets of annotated data, facial recognition accuracy has improved significantly in the last few years,” said Christopher Lam, Head of NEC Laboratories in Singapore.
Specifically worth mentioning is the fact facial recognition can maintain its accuracy even when people are wearing masks.
“The pandemic introduced a new dimension and imperative, with faces suddenly half-covered by masks. Face recognition suppliers rose to this challenge with an increased emphasis on the features of the periocular region around the eyes. Many algorithms now achieve the same accuracy rates for facial images with masks as for matching images without masks five years ago,” said Terry Hartmann, VP of Asia Pacific at Cognitec Systems.
Edge vs. server
Meanwhile, facial recognition algorithms have become lighter, allowing them to be run on the edge instead of on expensive servers. “Because facial recognition is a three-step process – that is, face detection, feature extraction and matching – the first two steps can actually be performed in the camera. So we are seeing some hybrid implementations where the camera performs the face detection, crops the face region and sends to the server for feature extraction and matching. This implementation is able to spread the load of facial recognition across the edge (camera) and the server, and typically can achieve better performance if done properly,” Lam said.
“The majority of enterprise-level facial recognition algorithms require powerful expensive servers which means that hardware costs prevent most customers from being able to implement facial recognition. The SAFR platform has the fastest and lightest algorithm which is able to run in camera or on a small low cost edge device without compromising accuracy,” Bennett said.
More diverse use cases
Facial recognition remains a premier security application in
various verticals. Yet more and more, it’s used to drive business intelligence and user experience, and the pandemic has accelerated this trend. “The past two years have seen the digital identity market growing at an exceptional rate. The pandemic situation has acted as an accelerator for many sectors to offer digital gateways for people to access business, government, travel and commerce services. Face recognition has become the key enabler for these digital identity schemes, especially for advancing mobile identity verification systems and remote onboarding technology,” Hartmann said.
Hartmann further cited some examples. “The travel industry has been the most forward thinking vertical … airports continue to invest in speeding up implementations of ‘seamless journey’ concepts to facilitate all transactions from curb to gate via face recognition and identification. The event industry is looking for touch-free, quick access control solutions to reduce interactions and fulfill various regulations. Clubs, pubs and event venues can quickly check in members and preregistered guests via face without looking at ID documents, thereby eliminating long lines at entrances and providing swift transactions without touching devices and surfaces,” he said.
Controversies
Despite its benefits, facial recognition also has issues
that cause worries and concerns for users. These are primarily in the areas of privacy, data collection and bias towards certain ethnic groups.
In terms of privacy, there have been reports of facial recognition systems using faces not only to identify individuals but also to unearth other data concerning those individuals from social media and the Internet. Good practices therefore are needed.
“With regard to face recognition technology in particular, in order to promote business activities based on the policy, in addition to developing internal processes to reduce risks related to human rights and privacy, we are working to ensure that the use of face recognition technology by our customers and partners, who provide our products and services, is an appropriate application that respects human rights. In addition to complying with relevant laws and regulations in each country and region, each and every employee gives top priority to respect for human rights at all stages of our business activities,” Lam said.
Closely related to privacy are controversies around the collection of data, which is used to train algorithms and should be well protected. “Regulators have found that unfortunately not all facial recognition companies do adhere to the laws in the region where they are deployed, which is why we have seen substantial fines and companies being forced to erase data from their datasets. At RealNetworks we have defined a set of guiding principles which transcend local laws which includes valuing those who use our product and are seen with our products and taking all necessary precautions to ensure individual rights will not be violated,” Bennett said.
Finally, bias towards certain ethnic groups happens from time to time due to poor training of algorithms, or training them with low-quality/mislabeled images. “Face recognition technology vendors have a responsibility to implement best practices that identify and minimize any hidden biases, establish metrics for fairness, and test algorithms in real-world scenarios. In recent years, the scientific community has been working together to improve training procedures, data and outcomes that reduce misidentifications not only based on gender, but also on age, ethnicity and other variables,” Hartmann said.