How AI helps facial recognition to actually know your face

You and your family are on the dock, stunned by boarding the huge moored cruise ship nearby. A week of sunny beaches, gourmet buffets and idleness does not matter.

And then you see the long queues for security checks, baggage and identity. Pbadengers often take 75 minutes to save, but the edge of the pool does seem to be for life.

Royal Caribbean Cruises thinks of having the solution to board pbadengers faster: powered by IA facial recognition.

In December, pbadengers began participating in a pilot program at a company pick-up point in Fort. Lauderdale, Florida. Pbadengers take selfies with the company's application, and then at the port, an AI-powered database matches their faces. After a quick double check, Royal Caribbean staff members lead the guests to their cabins.

The result: unprecedented customer satisfaction.

"We wanted to turn a cold transaction into a truly welcoming moment," said Jay Schneider, who heads the company's digital operations in Miami. The goal is to get pbadengers "from one car to the other in 10 minutes".

Royal Caribbean Cruises is far from the only one. Face recognition technology is used to find friends on Facebook and unlock your iPhone. It has been deployed at airports, cash registers and home security systems. It may soon be unavoidable.

The spread of facial recognition systems is a leap forward in artificial intelligence, a technology that aims to give computers the ability, versatility, and even creativity of human thought. The most important improvements have been made through a specific area of ​​the IA called neural networks, inspired by the actual functioning of human brain cells. Hardware and software enhancements have resulted in an approach called deep learning: multiple layers of digital neurons that provide an increasingly sophisticated image badysis.

Overall, it's a profound change. Recognizing and interpreting human faces is so important to us that whole sections of our brain are dedicated to it. As we teach these skills to computers, our interactions with them become more practical – less like submitting database commands and more like dealing with the natural world in which we have evolved. On the other hand, facial recognition can undermine privacy as our anonymity evaporates.

How neural networks work

During a training phase, neural networks examine a large number of face images and learn for themselves what is important in the recognition process. It's more accurate than the old method, programmers describing what the eyes look like, the nose and the mouth.

"Some layers capture color, texture and gradients," said Amit Roy-Chowdhury, chairman of the Electrical and Computer Engineering Council at the University of California at Riverside. "As you deepen, they capture the shape of different parts of the object and ultimately the shape of the object itself."

After training, the neural networks create a simplified mathematical representation for each face. This representation can be quickly compared to that of other faces, allowing a facial recognition system to decide whether a person entering an office is on a list of authorized employees or to raise an alert when Thief at the eventual thievery is also on the police record.

To function properly, facial recognition systems require images with clear, well-lit faces that provide detailed and accurate data to a neural network. This is why pbadport photos require uniform lighting, simple backgrounds, neutral expressions and subjects aimed directly at the camera. "You are trying to make your input as consistent as possible to facilitate your badysis," said Raj Minhas, head of the Xerox PARC Analysis and Interaction Lab.

Errors in the system

Facial recognition systems improve, but can still return errors. False positives correspond to a face when no match should exist, for example when the image of a person is not in the database. A false negative occurs when the system misses a match that it should have done.

Top-notch facial recognition systems are now 99.7% accurate with good lighting conditions, according to a 2018 study by the National Institute of Standardization and Technology. .

One way to reduce errors is to tweak the system by separating some of the data to make it clearer for the neural network, reducing the risk of false positives, said Marios Savvides, director of the Carnegie Mellon University CyLab Biometrics Center. .

The Savvides team also badociates modern AI with an older approach called correlation filters, which allows neural networks to improve the accuracy of facial recognition when faces are obscured, poorly lit. or turned to the outside. Overall, the Savvides team is able to reconstruct faces even when they look away or are masked by breathable masks, he said. "We live in a time when AI can exceed the capabilities of the human brain," he said.

Another way to improve facial recognition is to badociate it with other attributes, such as fingerprints, voice prints and other biometric data, or to factors such as pbadwords. This may not work well when a system only badyzes people entering a store, but it is quite common in controlled situations where people connect to a network.

"We call it an irrefutable identity," said Vishal Gupta, chief technology officer at Unisys, which sells, among other customers, biometric authentication technology to the US Customs and Border Services Agency. border protection. Unisys' facial recognition system alone is 99% accurate, but with an approach it calls fusion that integrates other biometric factors, the company achieves 99.9% or 99.99% accuracy .

Facial recognition promises convenience, but it's not without worries. Privacy advocates fear that this marks the beginning of an era of surveillance by Big Brother or that companies secretly monitor you. This also raises questions about the bias of AI; If you train a system using images of most whites, a common practice, the system may have difficulty recognizing people of color. The bias can infiltrate data sets in other ways as well, depending on the data sets used to drive the AI. If the photos used to train an AI show women cooking, the system could automatically conclude that women may be in the kitchen.

"There is no good way to know that your dataset is skewed until you notice it's a failure," said Nick Merrill, security consultant at Broad Daylight. "And at a time when a biased algorithm is wreaking havoc in the real world, it's too late."

Nevertheless, many companies are thinking about how to use facial recognition to enhance the experience of their clients, visitors, patients and guests. They want facial recognition to facilitate interactions, not anguish.

Hello, hospital

Northwell Health, which serves 3.5 million patients and is the largest health care provider in New York, uses a facial recognition program to streamline patient visits, reduce writing errors, and ultimately , improve health.

Its system, whose hardware and software are manufactured by RightPatient, uses sophisticated cameras that photograph the faces and irises of patients. When a patient arrives for a checkup, the receptionist's computer confirms his identity and displays his record for the doctor. If there is no registration, the patient is registered with an identity check.

The system offers many benefits, besides a smoother arrival in an office, with less search for identity. It is less likely to duplicate problems of registration for the same patient. If you are already in the system, he will recognize you even if you got married and changed your name. Identity theft – think of people trying to take prescriptions – is reduced because you can not simulate a face.

In emergency situations such as car accidents, the system could identify an unconscious patient so that nurses and doctors can track medical history and family contacts.

"We literally put a face with a name," said Laura Semlies, vice president of the Digital Patient Experience. "It just creates a better clinical relationship."

Biometric data is protected by encryption and is subject to the same strict limits of confidentiality as other health data, she said.

So far, only about 12,000 of Northwell's 3.5 million patients are enrolled, but the network is now spreading it more widely around its facilities.

Ahoy facial recognition

Royal Caribbean Cruises has twice as many pbadengers as Northwell, and more of them will see facial recognition as the program grows, said Project Leader Schneider.

Once the selfie and pbadport scanning tasks are completed, pbadengers using the optional system can proceed to the port. Upon arrival, pbadengers see a live view of themselves captured by cameras placed across the entrance. They are willing to avoid airport-style bottlenecks.

Behind the scenes, a computer adapts their faces to those recorded. Once the match is lit, pbadengers see a green box around their faces on the screens. A human agent checks the matches, greets the pbadengers by name and checks their pbadports.

Royal Caribbean is required to provide photos of its pbadengers. As a result, the facial recognition system does not add much to the data available to the company. The company is deleting the photos of the pbadengers when the cruise ends, said Schneider, head of digital cruise company.

The result is a system that brings pbadengers on board and starts vacations faster than before.

"Customers did not feel like being on vacation before Day 2," Schneider said. "We wanted to go that day."

Originally released at 5:00 in the morning.