- Facial recognition technology can now identify disguised faces
- Large stores of training data required, taxing traditional computing infrastructures
- Some see the technology as a privacy intrusion — others see enhanced security
We live in an age of video surveillance. Like it or not, going out in public means being caught on someone’s security camera. It’s a creepy fact of modern life.
If you’re a law-abiding citizen, you may (or may not) have a problem with this intrusion of your privacy. Criminals, on the other hand, are not so happy to have the eyes of the world on them.
Researchers began developing manual face recognition systems in the early 1960s. Automated face recognition emerged at the beginning of the 1990s. Today, with the aid of big data and artificial intelligence, face recognition technology has a variety of uses including verifying passports at airports, employee clock-in, and violent crime prevention.
Of course, as the technology improves, criminals will try to outsmart it. They might wear wigs, glasses, or false beards to mask their identity. Amarjot Singh at the University of Cambridge is working on a way to outfox the foxes. Singh and his colleagues at the National Institute of Technology in Warangal, India and the Indian Institute of Science in Bangalore have developed a deep learning method called the Disguised Face Identification (DFI) Framework.
Prior to Singh’s work, disguised face recognition was attempted using color and roughness of skin as identifiers. This system was ineffective because when disguised, a large part of the face is not visible.
Singh’s method detects 14 facial key-points using the Spatial Fusion Convolutional Network. The key-points are marked manually by a human on images of disguised faces. A deep learning network then examines a batch of disguised face images and learns to predict the key-points, hidden or otherwise.
Facial recognition technology is coming of age. ~ Amarjot Singh
The network then tries to correct the prediction error made for each key-point until it can predict them reasonably well. The key-points are connected to form a star net structure that is different for every individual face.
DFI is a challenge for scientists because it requires a large amount of annotated training data and the existing databases appropriate for this type of study are too small. To overcome the problem of limited data, Singh’s team created a custom dataset of 2,000 images of people ages 18 to 30 wearing different types of facial disguises in front of various backgrounds and lighting schemes.
The researchers performed experiments on one dataset with disguised faces on simple backgrounds and another on a second dataset with disguised faces on complex backgrounds. The team found that the framework captured facial key-points more accurately on images with simple backgrounds.
Key-point detection in the eyes, nose, and lip regions of the face was of particular interest to the researchers. The framework detected points in the nose region with similar accuracy when examining images in both datasets. Point detection in the outer areas of the eyes and lips was less accurate for faces on complex backgrounds.
When applied to images containing multiple faces, the framework’s accuracy decreased significantly.
Accuracy also decreased as the complexity of the disguise increased. The DFI outperformed existing state-of-the-art face recognition technology, albeit marginally, which indicates Singh’s work is a step in the right direction.
Critics worry that disguised face recognition technology could lead to human rights abuses, but Singh believes his system will help to tackle global security challenges by identifying masked criminals that pose serious threats.
“The new iPhone can be unlocked simply by looking at it, and accessing your smartphone is just one of many ways that facial recognition will change our daily lives,” notes Singh. “Facial recognition technology is coming of age.”