Anil Jain is fingerprinting infants, but not because of a baby crime spree.
Instead, the Michigan State University professor wants to use GPU-powered artificial intelligence and fingerprints to help keep kids healthy.
Many children in the developing world lack any form of identification, making it hard to know which babies have received which vaccine or booster shot.
“We want to make sure there’s complete coverage, that every child is given the right vaccinations,” Jain said. “With fingerprint-based health records, we can track this.”
Fingerprints could also be used to identify missing children or resolve cases in which newborns are swapped at birth, Jain said.
Tiny Fingerprints a Challenge
Capturing a usable fingerprint from the soft, pliant skin of often-fidgety babies has been just about impossible. That’s especially true for infants at four weeks, the age children receive their first vaccination.
“When you put a finger on the sensor and apply the slightest pressure, there’s a lot of distortion in the fingerprint image,” Jain said. The ridges and valleys on the fingerprint are not yet well-defined, so the contrast is poor.
Jain and his team developed machine-learning algorithms to enhance the quality of fingerprint scans. They used NVIDIA Tesla K40 GPU accelerators to train their deep neural networks and GeForce GTX TITAN X GPUs for testing in the field.
They also worked with fingerprint-scanner maker NEC to create a scanner designed for infants, with more than twice the resolution of a standard scanner. Standard scanners work for older children but are not adequate for infants.
Anil Jain and his team fingerprinted infants during four visits to Agra, India.
Deep Learning Fingerprint Matching
After fingerprinting more than 300 babies at Saran Ashram Hospital in Agra, India, Jain and his team were able to identify infants who were first fingerprinted at 6 months or older with nearly 99 percent accuracy. But accuracy rates fell to 80 percent for babies fingerprinted at four weeks.
The researchers knew they had to raise accuracy rates to put their work into practice. “That’s why we’re designing our own fingerprint-matching (software) and training it using GPUs and deep learning,” Jain said.
Previously, Jain used commercial software to match one set of a child’s fingerprints to the other. Now his team is using the CUDA 7.5 programming model and TITAN X GPUs to train its neural network to recognize and match fingerprints.
Fingerprint-based Healthcare Record
When accuracy improves, Jain hopes to see the technology used in hospitals, by healthcare workers who serve several villages and in the small, sometimes remote clinics where many children receive health care.
Although some have raised concerns about fingerprinting children, Jain said he believes the risks, if any, are minimal, and outweighed by the potential benefits.
“If a child isn’t vaccinated, you can’t eliminate disease,” Jain said.
For more information about the team’s work, see their paper, “Giving Infants an Identity: Fingerprint Sensing and Recognition.” Image and video courtesy of Michigan State University.