August 08, 2014
According to the University of Central Florida (UCF), graduate student Afshin Dehghan and colleagues from the UCF Center for Research in Computer Vision have created a facial recognition tool they think is capable of matching pictures of children with their parents.
Dehghan graduated from the University of Tehran in 2010 with a bachelor’s degree in applied science. He then came to UCF where he expects to get his doctorate next year.
The project started with more than 10,000 online images of famous parents and their offspring.
“We wanted to see whether a machine could answer questions, such as ‘Do children resemble their parents?’ ‘Do children resemble one parent more than another?’ and ‘What parts of the face are more genetically inspired?’” Dehghan explained.
While these topics are commonly discussed by anthropologists, the UCF team is looking to inspire new technological advances that can objectively analyze and evaluate data without human bias.
Ross Wolf, the associate dean in the UCF’s college of health and public affairs, as well as an associate professor of criminal justice and a law enforcement officer for more than two decades, said the tool could potentially be “used to identify long-time missing children as they mature.”
While law enforcement officials already use facial recognition software, those programs lack the ability to identify the same features and characteristics in photos over time. The newly developed program is intended to do that.
Using actress Catherine Zeta Jones and her children as an example, they set out to use the technology to identify lesser-known features that the relatives have in common.
While people might look for something prominent, such as a shared smile in the Hollywood star and her offspring, the researchers were hoping to find shared traits not necessarily viewed as significant, such as the chin, parts of the forehead, and even the left eye. They developed an algorithm to focus on those features, converting photos into a checkerboard of small regions and pulling out small snapshots of various segments of the face.
“The computer compared all the photos feature by feature and sorted them by the most probable match,” the university said. The goal was to see if the computer could correctly link each child with its parents. “The team found that its program not only did a better job of matching features of parents and their kids than random chance, but it also outperformed existing software … by 3 to 10 percent.”
Furthermore, the research provides evidence that children resemble their parents, often in unseen ways. In 63 percent of the cases, sons were more likely to resemble their fathers more closely than their mothers, while 82 percent of daughters were more likely to resemble their mothers, the authors said.
“Machines can learn through time. When a computer goes through thousands of images it knows what it has seen and is able to tell you,” said Dehghan.