Startup Xjera Labs develops advanced facial recognition system
Singapore-based startup Xjera Labs has developed a face recognition system that uses artificial intelligence to detect and identify a person or an object from CCTV footage within minutes, according to a report by International Business Times.
In development for the past five years, Xjera Labs’ neural networks can detect individuals, vehicles and objects from video footage with a 97 percent accuracy, based on a database of 20,000 facial images taken from CCTV footage.
Neural networks are large networks comprised of AI classical computers that are trained using computer algorithms to solve complex issues, whereby certain layers analyze parts of the problem and accumulate to generate an answer.
Despite the extreme difficulty in spotting an individual in a crowd, even when reviewing video footage, Xjera Labs has managed to pull off this feat by building ultra dense 52-layer neural networks and using multiple networks.
As a result, the system can answer a single attribute question such as, “How many men have brown hair?” within 200 milliseconds.
The user is required to run several search requests to narrow down the potential answers, expediting the search process to take only a few minutes rather than hours or days.
Breaking down a face into a multitude of attributes
“If the police want to find someone, then they want a very fast response,” said Ethan Chu, founder and CTO of Xjera Labs. “With our system, we are constantly doing facial indexing in real-time on CCTV camera footage. The footage is uploaded to the cloud by our customers and we process the data.
“We use part-based representation and different layers in the neural network focus on different attributes. We have basic layers that describe the subject in general, ie. the subject’s shape, texture and colors. Then other layers are split into different attributes to describe things in more detail, so if we want to know if someone is wearing glasses, only the layers that concern detecting the head will be used to search for that.”
Xjera Labs currently has three products including XHound, which can locate a person or vehicle of interest; XIntelligence, which is able to determine the number of people in a crowd in both indoor and outdoor settings; and XTransport, which can determine the number and make of vehicles on highways, as well as detect illegal driving and traffic accidents.
All three systems run on six neural networks that each have a large number of layers. The first network detects the individual’s mood by analyzing their facial expression; the second detects and recognizes actions, the third focuses on facial attributes; the fourth distinguishes people from the background in the video; the fifth detects text in the form of licence plates, signs, logos, alphabet letters and language characters; and the sixth categorizes vehicles by their make and type.
“We use deep learning algorithms and our own neural network architecture that we started developing in 2012,” said Chu. “It utilizes very few GPU resources. We worked together with Nvidia and just one P4 Nvidia GPU can support 32 cameras in real time concurrently and filming in HD.”
To ensure that the database’s millions of facial images is not compromised and exploited by criminals, Xjera Labs makes it a point not to store any of the images.
Rather, the company uses a technology it developed called “feature transformation” in which the system extracts details about an individual’s physical features — such as whether or not the person wears glasses, their height, sex, race, hair color — and encrypts the information.
If an individual commits a crime in Singapore and is captured by CCTV cameras, the facial recognition system will assign their identity to a series of numbers.
The police can then set up an alert that will notify them if the criminal is ever captured by another CCTV camera in any part of the country — even if it’s five years later.
Xjera Labs products are currently being used by the Singapore police and schools in China.