Deniz Aykac takes a break from collecting images for a large data set of faces and bodies that can be used to train machine learning software to recognize individuals. Credit: Carlos Jones/ORNL, U.S. Dept. of Energy
Deniz Aykac knows your facial expression when you’re texting while you drive. She has made an atlas of your lungs and used neutron-vision to peer inside sealed containers. Now she is teaching your car to see.
A very specific kind of visionary, Aykac is an imaging expert. She characterizes and processes information from optical, infrared and other kinds of images. Unlike some researchers who focus on a specific discipline for an entire career, she uses her skills to advance a variety of different scientific fields, constantly switching gears and roles.
Today, Aykac is designing experiments to test software for the Biometric Recognition at Altitude and Range project, or BRIAR, which is honing long-distance and aerial recognition to fight terrorism and protect critical infrastructure. Aykac helped collect video of volunteers to create the largest data set of its kind for teaching computers to identify the faces, bodies and walking gait of individuals seen from long-range cameras and elevated angles. Aykac helped curate and arrange the videos, then compare the performance of recognition software designed by seven different teams.
Long before coming to the Department of Energy’s Oak Ridge National Laboratory two decades ago, Aykac had learned to visualize her own winding path forward—even when it seemed invisible. Her unconventional route to an unusual scientific career in Oak Ridge required her to learn on the fly, overcome obstacles and let her curiosity be her guide.
Her physics career was determined for her by a 3.5-hour exam in her native Turkey. Unfortunately, it turns out she’s not wild about theoretical physics. Aykac was assigned to a college where all the lectures were in English, which she didn’t yet speak. She rapidly learned so she could attend Boğaziçi University, despite its limited resources.
After that, she pursued a master’s degree in applied physics at the University of Iowa. But her enthusiasm for a class in biomedical image processing led her to switch her academic focus. Although her college education had left Aykac underprepared for this field, she was determined. Her first big project involved segmenting images of airways for a lung atlas.
“With almost every project there’s always a learning curve,” Aykac said. “I had to make up for that time I didn’t have the biomedical undergraduate degree. I had to teach myself a lot of extra hours to come up to speed.”
At the time, she recalls, “Image processing was all about coding.” This was not good news for a student whose undergraduate lab had four computers. Without a single coding course, Aykac started learning to do it on her own at age 23. Nevertheless, she relished the challenge: “You can do things that give you so much satisfaction,” she said.
Even when Aykac seemed to hit a dead end, she continued to forge her professional path. After moving to Knoxville with her husband, she applied for an ORNL job originally geared toward a candidate with a doctorate. When that hire didn’t pan out, Aykac earned her way in through her work on a National Institutes of Health project: As a postmaster’s research associate, she made a good impression developing a 3D analysis tool. It was able to isolate images of organs in CT scans of mice. Aykac was offered a part-time position in the imaging, signals and machine learning group—which she parlayed into a career.
Her research scope rapidly expanded beyond biology and medicine. Aykac developed methods and software for inspecting coated particle nuclear fuel; implemented detection codes for malfunctioning radiation detectors; and used fast-neutron transmission tomographic images to understand the size and shape of potentially radioactive contents inside closed containers.
One of Aykac’s career highlights came as a member of a team that created a tool to automate eye screenings for diagnosing blinding diseases, especially for diabetic patients. The project was recognized with an R&D 100 Award in 2010 as well as the UT-Battelle Award for Excellence in Technology Transfer.
For almost a decade, Aykac has been involved with image processing projects to better recognize links between driver behavior, road features and safety. Some of the data sets she developed can be used as a baseline for improving the design and performance of self-driving cars.
For example, several naturalistic driving studies relied on visual analysis of driver faces and actions. Volunteers on the ORNL campus sat in the driver’s seat of a stationary car while Aykac asked them to perform different actions, like singing or fiddling with the radio, as cameras recorded them from different angles.
“It’s a fully-annotated public data set, which is really valuable for people who are working on any facial recognition, emotion recognition or activity,” said Aykac, who also provided the annotation. “It could also be used for training machine learning to recognize what type of distraction might have led to a crash.”
Sometimes Aykac follows a line of research on a side trip. The transportation studies routed her to rumble strips. Front-view camera footage from cars had been collected, and Aykac estimated the rumble strip size for use in software analysis. Aykac converted the road video to homography images, resembling an overhead shot looking down on the car in its lane. The rumble strip data could be incorporated with other footage to train self-driving cars to avoid hazards.
Simultaneously, Aykac kept her hand in biological studies. She recently completed a project using optical imaging to help define how microbes living together affect each other’s growth. She also advised the ORNL campus greenhouse as it developed a system to analyze plant health using a series of cameras with various capabilities.
Aykac enjoys being able to shift gears among different types of projects. “It’s a nice break for me to reset,” she said. “Then when I go back, I can see the things I missed. Something gets cleared.”
This spring, Aykac will be involved in testing which software developed through BRIAR is best at recognizing individuals in challenging conditions, such as when the video quality is degraded or when a person’s body is visible but not the face.
Aykac’s experiences laid the groundwork for her to frequently shift angles, learn new skills and recognize new ways to apply them. This is one of the reasons she enjoys working at ORNL.
“I like the flexibility. I think the challenges make me a better researcher,” she said. “Some subjects are so appealing to me that I want to learn more, following threads. I never lost that.”
UT-Battelle manages Oak Ridge National Laboratory for the U.S. Department of Energy’s Office of Science. The single largest supporter of basic research in the physical sciences in the United States, the Office of Science is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science.
