October 06, 2021
Mizzou Engineers have developed a two-stage computational model that can accurately predict someone’s age based off a single facial image. The system earned the team third place at the Guess the Age contest, part of the biannual 19th International Conference on Computer Analysis of Images and Patterns held in Cyprus.
Finishing third is especially noteworthy considering it was the first time the research team — from the Computational Imaging & VisAnalysis (CIVA) lab — has applied their video analytics expertise to facial recognition, said CIVA Director Kannappan Palaniappan, professor of electrical engineering and computer science. CIVA researchers focus on computer vision and video processing algorithms to analyze aerial and biomedical imagery.
Palaniappan also noted that the team was up against heavy hitters in the industry. The winning team, for instance, included scholars from the Chinese Academy of Sciences and researchers from Baidu, one of the largest Internet companies in the world.
“I was extremely pleased with how well our team was able to do in a new application area within a short time window,” Palaniappan said.
For the competition, teams were given more than 575,000 photos of celebrities from various sources, such as publicity shots and stills from video clips from the Oxford University VGGFace2 dataset with age annotations added by the challenge organizers at the University of Salerno in Italy. This is one of the largest databases of facial images with age labels. Teams used that data to train artificial intelligence learning models to associate image features with corresponding ages. The submitted algorithms were tested using a sequestered private set of almost 170,000 photos for evaluating accuracy.
The CIVA team used a novel two-stage deep learning architecture with a ResNeXt backbone plus random forest regression to pinpoint the age.
“The AI model we created was a combination of deep learning and traditional machine learning methods,” said Imad Toubal, a PhD student in computer science. “The deep learning stage processes an image through different layers and converts the facial image to a high-dimensional vector representation, like a numeric-fingerprint, for that face. The vector embedding is then used in a cascading sequence of random forests where the first step predicts the age group of someone; for instance, the likelihood of a person being between the ages of 11 and 20. The second random forest step is more specific and predicts the actual age rounded to the nearest year.”
In most cases, the system was able to correctly estimate ages within two years of their true age, even when faces were partially covered with glasses, hats, facial hair or cosmetics. Mizzou’s results were on par with results from the top two teams. The AI model works across gender and demographics, and is robust to variations in photo quality, lighting and head pose.
Having a machine determine someone’s age based on an image has numerous applications, from helping law enforcement identify and search for missing persons to protecting minors by restricting inappropriate websites.
It could also help protect privacy by eliminating the need to reveal personal information when trying to show proof of age. For instance, a driver’s license can confirm someone’s age with the date of birth, but it also includes a license number, home address and physical traits such as height and weight.
“There are consumer applications where you may need to determine the age of a person before they can purchase a product or enter a venue,” Palaniappan said. “Our AI software would do that without a person having to show ID. Future work will be to determine how this can be used to enhance and improve differential privacy in social media that is location-dependent, which is part of our joint National Science Foundation project with Dan Lin at MU and Chris Clifton at Purdue for facial image protection with provable privacy guarantees.”
Lin, an associate professor in EECS, and graduate student Linquan Lyu were coauthors on the paper (Single View Facial Age Estimation Using Deep Learning with Cascaded Random Forest) describing the face estimation algorithm.