Algorithms for Facial Recognition

Various facial recognition technologies are used on a daily basis—from simple tasks such as tagging someone in a photo on Facebook to more complex undertakings like tracking down a suspect in an FBI investigation. There is an undeniable demand within the intelligence community for a high-speed system that can also deliver accurate results. In most cases, the photos being processed are not straightforward headshots, which significantly slows down facial analysis. Led by Dr. Melissa Smith in the Department of Electrical and Computer Engineering, a group of computer engineering students in the Future Computing Technologies Creative Inquiry uses high performance computing to create a more efficient facial recognition system. First and foremost, the team must identify troublesome characteristics of the current facial recognition systems. “There are all kinds of variables that you have to account for, such as an unusual angle or dim lighting,” senior Ben Shealy said. Even a beard or a shadow on the face can thwart the system; the team is faced with the challenge of finding techniques that account for those weaknesses.

There are a multitude of algorithms that can be used to process this type of data. “We want to take a current picture of the person, apply several machine learning techniques that each have different strengths and weaknesses and then intelligently interpret the output and make a final classification decision,” Jesse Tetreault, a computer engineering graduate student working with the team, said. To narrow it down, the team opted to use three of the most popular algorithms in facial recognition systems. They found that these three algorithms are very efficient when used in conjunction with one another. Students primarily work on laptops in the lab to try to achieve quicker local computation. As the project progresses, the team begins taking advantage of high-performance computing technologies like the Palmetto Cluster supercomputer, which has graphic processing units that can run different algorithms in parallel.

While working carefully to finalize the three algorithms, the team frequently checks the accuracy of their computations. The team uses a public database with 400 photos of 40 different people. Each person was photographed ten times, from different angles in different lighting, enabling the team to determine which algorithms work best in the shortest amount of time. Increasingly, this Creative Inquiry project is getting closer to their central goal of local computation in real time so that facial recognition technologies can become more efficient and accurate.