Facial recognition is more than 50 years old.
A research team led by Woodrow W Bledsoe ran experiments between 1964 and 1966 to see whether ‘programming computers’ could recognize human faces.
The team used a rudimentary scanner to map the person’s hairline, eyes, and nose. The task of the computer was to find matches.
It wasn’t successful.
Bledsoe said: “The face recognition problem is made difficult by the great variability in head rotation and tilt, lighting intensity and angle, facial expression, aging, etc.”
Computers find it harder to recognize faces than beat Grandmasters at chess. It would be many years before these problems were overcome.
Thanks to camera technology improvements, mapping processes, machine learning, and processing speeds, facial recognition has come of age.
Most systems use 2D camera technology, which creates a flat image of a face, and maps ‘nodal points’ (size/shape of eyes, nose, cheekbones, etc.). The system then calculates the nodes’ relative position and converts the data into a numerical code. The recognition algorithms search a stored database of faces for a match.
2D technology works well in stable, well-lit conditions such as passport control. But it is less effective in darker spaces and cannot deliver good results when the subjects move around. It is easy to spoof with a photograph.
One way to overcome these flaws is with liveness detection.
These systems will look for indicators of a non-live image, such as inconsistent features between foreground and background.
They may ask the user to blink or move. They are needed to defeat criminals who try to cheat facial recognition systems by using photographs or masks.
Another critical advance is the ‘deep convolutional neural network.’
This is a type of machine learning in which a model finds patterns in image data.
It deploys a network of artificial neurons that imitates the functioning of the human brain.
In effect, the network behaves like a black box.
It is given input values whose results are not yet known. It then makes checks to ensure the network is producing the expected result. When this is not the case, the system makes adjustments until it is correctly configured and can systematically produce the expected outputs.
Today, previously advanced processes are finding their way into mass-market devices.
For example, Apple uses 3D camera tech to power the thermal infrared-based Face ID feature in its iPhone X. Thermal IR imagery maps the patterns of faces derived primarily from the pattern of superficial blood vessels under the skin.
Apple also sends the captured face pattern to a ‘secure enclave’ in the device. This ensures the authentication happens locally and that the patterns are not accessible by Apple.
Measurements and accuracy
Three criteria assess facial recognition systems.
1. False-positive (aka false acceptance)
This describes when a system erroneously makes an incorrect match. The number should be as low as possible.
2. False-negative (aka false rejection)
With a false negative, a genuine user is not matched to his or her profile. This number should also below.
3. True positive
This describes when an enrolled user is correctly matched to his or her profile. This number should be high.
These three measurements are conveyed in percentages. So, let’s say an entry system assesses 1,000 people a day. If five non-approved people are allowed in, the false positive rate is five in 1,000. That’s one in 200 or 0.5%.
So, what percentages do the current systems achieve?
The National Institute of Standards and Technology (NIST) regularly tests multiple systems to search a database of 26.6 million photos.
Its 2018 test found that just 0.2% of searches failed to match the correct image, compared with a 4% failure rate in 2014. That’s a 20x improvement over four years.
NIST computer scientist Patrick Grother says: “The accuracy gains stem from the integration, or complete replacement, of prior approaches with those based on deep convolutional neural networks. As such, face recognition has undergone an industrial revolution.”
Liveness detection systems will look for indicators of a non-live image such as inconsistent features between foreground and background. They may ask the user to blink or move
Further confirmation of the tech improvement came in the Department of Homeland Security’s Biometric Technology Rally in 2018. In its test, Gemalto’s Live Face Identification System (LFIS) scored a 99.44% acquisition rate in under five seconds, compared to the average of 65%.
Facial recognition vs. face detection: a significant difference
Though ‘facial recognition’ is generally used as a catch-all term, this is not entirely accurate. There is a crucial distinction between facial recognition and face detection.
Facial recognition describes the process of scanning a face and then matching it to the same person on a database. This is the approach used to unlock phones or authenticate a person entering a building.
Face detection is when a system tries to establish that a face is present. Social media companies use face detection to filter and organize images in large catalogs of photos.
The tools used to train the two systems are different. The desired levels of accuracy vary too. Facial recognition for identification purposes needs to score more highly than any system used to organize images merely.
The confusion between the two processes has caused some controversy.
In 2019, a researcher revealed that Amazon’s systems were much better at classifying the gender of light-skinned men than dark-skinned women.
This led to fears that surveillance systems could make more false matches for some ethnic groups.