The Sensors Exploitation Research Group at the Air Force Institute of Technology, led by Dr. Michael J. Mendenhall, has honed in on the process of differentiating human skin from other materials within an image to reduce false detection.
Color-image based systems, useful in locating people injured or missing in aerial search, rescue, and recovery operations and also for security and surveillance, have high false detection rates making it difficult to locate people in the collected imagery. Instead of using bulky, expensive and relatively slow hyperspectral camera systems, Mendenhall’s research team has developed a prototype camera system that specifically works with a skin detection and color estimation approach. The system requires only a small number of spectral channels.
Mendenhall, an Assistant Professor of Computer Engineering at AFIT’s Department of Electrical and Computer Engineering, along with his research team use a multispectral camera system to enhance skin detection by focusing on the amount of melanin in the skin.
Melanin is the primary component in determining skin color.
“Our approach concentrates on the melanin and water in skin. It can detect the skin while providing a means to determine how much melanin it contains. Since melanin is the primary element contributing to skin’s color, it’s a valuable piece of information to extract,” said Mendenhall. “I can use our camera system to filter out skin types based on the details of the person of interest. We can show only fair skinned people, only dark skinned people, or anything in between. This is particularly useful in speeding up the search process and improving an analyst’s ability to locate persons of interest.”
Mendenhall’s system is capable of real-time detection and color estimation at typical video speeds, at about 10 percent of the foot print and 10 percent of the cost. The system combines the necessary information from the visible and near-infrared light spectrum so that it enables them to distinguish between human skin and other common material such as grass, trees, and building tops, while providing the ability to filter out skin colors that aren’t of interest.
Currently, Mendenhall’s team is looking to improve the performance of the skin detection approach by incorporating the mirror-like reflection off of skin into their algorithms. In the future, they would like to account for the hair on the skin, so they can make additional improvements in estimating skin color.