Research shows gains in biometric identification from thermal images as contracts increase

A scene in the science-fiction film The Matrix Reloaded shows the simulated thermal image of actor Hugo Weaving’s angry face. The blazing, shifting and blurry image was intended to make Weaving’s character look monstrous but still recognizable.

In reality, thermal imaging obscures identifying features enough that some hospitals use it to monitor patients’ safety while in restrooms without needlessly invading their privacy. Now, civilian and military researchers are looking for affordable and reliable ways to identify people caught on thermal imaging.

Just this month, Intel Corp. researchers said they have used facial recognition software that is trained to identify faces in visible light to identify faces recorded by thermal cameras. Significantly, they did not have to retrain the model to do so.

Also in January, a news report surfaced indicating that the U.S. Department of Defense is building deployable systems that can identify people in thermal images.

In Intel’s case, scientists worked with a FaceNet deep neural network architecture. The company says that deep neural networks have gained prominence “due to their human-like competence” for person recognition using visible light only. Researchers used a model trained on VGGFace2, a large-scale face recognition database introduced in 2018.

In their experiment, scientists used two datasets. One, created by Intel researchers using a FLIR ThermaCAM* SC3000 camera, contained 766 images of 40 volunteers (21 women and 19 men). The second was a public dataset from Oklahoma State University containing 4,190 thermal images.

Using Intel’s software model, the company’s researchers distinguished their 40 volunteers 99.5 percent of the time. Volunteers in the Oklahoma State dataset were distinguished 82 percent of the time.

Meanwhile, an article in OneZero, a Medium technology and science publication, reports that thermal-image contracts running from Sept.2019 to 2021 valued at $4.5 million have been announced by the Pentagon.

The military wants nothing less than a service-member-portable sensor capable of identifying individuals from thermal images recorded through automotive windshield glass, fouled by foggy conditions or marred by strong backlight. The camera needs to have a 10- to 500-meter range, and the accompanying software must match subjects against a watch list.

Cyan Systems Inc., in Santa Barbara,Calif., won a $2.8 million development contract. Cyan has a patent on related technology.

Polaris Sensor Technologies Inc. of Huntsville, Ala., was awarded a separate $1.5 million contract to work on a similar development project. Polaris also has a previously awarded patent for a method of increasing details in thermal images.

Article Topics

biometric identification  |  biometrics  |  FaceNet  |  FLIR  |  identity verification  |  Intel  |  military  |  research and development  |  thermal imaging