Facial recognition performed with thermal camera and neural network
Image credit: Dreamstime
The US Army Research Laboratory has developed a facial-recognition system capable of identifying faces from thermal images. This could pave the way for real-time and post-mission forensic analysis to assist with covert night-time operations.
Thermal and night-vision goggles detect electromagnetic radiation of lower frequency than visible light (infrared); with the latter, in additional to visible light. This allows operators to view action and objects - such as humans and electrical equipment - based on the heat signatures that they emit. Thermal cameras are used for surveillance in check points and watch towers and increasingly on body cams.
The ability to perform facial recognition using images from thermal cameras would enhance this surveillance, such as by automatically detecting when a person of interest is nearby and sending an alert to operators.
“This technology enables matching between thermal face images and existing biometric face databases/watch lists that only contain visible face imagery,” said Dr Benjamin Riggan, a research scientist who was involved in the development of the technology.
“The technology provides a way for a human to visually compare visible and thermal facial imagery through thermal-to-visible face synthesis.”
According to Riggen, under night-time and low-light conditions, a conventional camera - which captures visible light - is not able to capture images without use of an attention-grabbing flash. This would immediately reveal the location of the camera.
While thermal cameras do not have this drawback, matching a thermal image of a face with an image of a face captured with a conventional camera is a challenge known as ‘cross-spectrum’ facial recognition.
The US Army researchers approached the challenge using techniques based on deep neural networks: multi-layered computer programs which loosely mimic the structure of biological brains, often used for image analysis. They employed a non-linear regression model to map a thermal image into a representation in visible light and an optimisation problem which preserves the shape of the whole face such that a synthesised visible image can be generated.
The researchers were able to perform facial verification using this approach and found that it was effective. It was capable of performing recognition almost in real time while running on a laptop.
According to Riggan, the group will continue the research and aim to develop the technology for military use.