Researchers at the Karlsruhe Institute of Technology in Germany are working on a deep machine learning system that can match the faces from infrared cameras to their visible light counterparts. Rainer Stiefelhagen and Saquib Sarfraz say that while the system they’re developing isn’t yet perfect, it marks a major improvement over other such systems.
The difficulty of matching these different kinds of images lies in how they capture light: While standard visible light cameras capture the standard spectrum reflected off of their subjects, infrared cameras focus on light emitted by their subjects, which varies by factors like body temperature, subjects’ activity, and so on. Essentially, what the researchers have done is train software to learn how the images from a database of visible light and infrared light pictures compare with each other. As detailed in an MIT Technology Review article, the “neural network” system continually scanned through thousands of these images, looking for patterns in how they compared and ultimately developing an ability to match faces between the different kinds of images.
While the researchers say their system is 10 percent more effective than other leading technologies attempting the same task, its accuracy still lags at 80 percent when it has numerous visible light images to compare against, and 55 percent in one-to-one comparisons. Still, as it develops, it could turn into a useful tool, particularly in the area of law enforcement. Police around the world are already scanning CCTV and security camera footage with facial recognition technology to identify suspected criminals, and adding the capability of doing this with infrared images can only improve their results; it could also be useful in real-time security and surveillance monitoring. If Stiefelhagen and Sarfraz’s neural network system can learn fast enough, we could soon see such applications of this technology in the field.
Source: MIT Technology Review
July 28, 2015 – by Alex Perala