TechTheDay

Current facial recognition systems are based on the matching of photos taken under generous lighting. These photos have good amounts of details with minimal noise and distortions. If the quality of these photos are reduced due to poor lighting, it’s likely that facial recognition systems lose their accuracy or refuse to work altogether. Unfortunately, not every place where facial recognition systems are deployed are well-lit.

Light is one of the most important factors in facial recognition. Without it, even state-of-the-art imaging hardware and software will have extreme difficulty recognizing faces. However, thanks to the work of a group of German researchers, it is now possible to do facial recognition in poor lighting conditions or even in complete darkness.

Thermal Imaging Based Facial Recognition

A group of German scientists is claiming that they already have the solution to this problem in facial recognition. They have developed a system that uses thermal signatures to facilitate the recognition of faces under unfavorable lighting conditions or even in complete darkness. This system analyzes mid-infrared or far-infrared images and compares them with standard photos. Basically, this system simply augments the capabilities of standard facial recognition. The use of infrared light or thermal imaging does not supplant the use of light in the visible spectrum. Both modes are used.

The details of this new facial recognition tech is detailed in the paper entitled “Deep Perceptual Mapping for Thermal to Visible Face Recognition” published on Cornell University’s ArXiv under the Computer Vision and Pattern Recognition category. M. Saquib Sarfraz and Rainer Stiefelhagen of Karlsruhe Institute of Technology in Germany are listed as the authors.

Reconciling the Images Taken Under the Thermal and Visible Spectrums

You would likely be thinking how come this idea has not been explored much earlier. It’s obviously a simple concept. Many should have already thought about it. The thing is reconciling the images produced under visible light and under infrared (non-visible) is not easy. The researchers considered thermal-to-visible face recognition as the most challenging problem.  They needed to bridge the two modalities while overcoming the highly non-linear relationship between the two.

Images taken under visible light are typically consistent and identifiable. It’s just a matter of properly capturing the details. The images produced under the illumination of visible light have common and constant attributes such that even when the visible light is decreased or increased, unique individual faces can still be distinguished and identified. In the case of thermal imaging, the resulting images keep varying due to a number of factors including the temperature of the surroundings, the temperature of the skin, and the physical activity and expression of a person. Moreover, the images taken using infrared cameras have significantly lower resolutions.

So how did the researchers reconcile the differences? They had to employ a deep neural network system.

Deep Neural Network System

A deep neural network is an artificial neural network designed to model complex nonlinear relationships. Basically, what it does is to imitate the brain’s ability to learn and identify objects. It has been used in language modeling and acoustic modeling for automatic speech recognition. Google similarly employed a deep neural network system to help counter spam.

Through a deep neural network system, the researchers have created a facial recognition tech that is capable of making connections and drawing conclusions based on the complex of factors under observation.

A deep neural network requires data from which it can establish connections and draw conclusions. For the initial study, the researchers fed their deep neural network with the University of Notre Dame data set, which consists of a considerable amount of standard and thermal (infrared) images of people. This data set contains 4,585 standard and infrared images of 82 people.

Proving that the System Works

With the systems ready, the researchers tested their new facial recognition technology by separating the images of the 82 people into two groups. The first group (images of 41 people) was used to rain the deep neural network system while the second group was used to test it.

The researchers found that the facial recognition system they created outdoes its existing counterparts. According to Sarfraz and Stiefelhagen, their approach improves the state-of-the-art by more than 10%. The system is reportedly capable of matching images and recognizing faces within 35 milliseconds. The researchers note that it is “very fast and capable of running in real-time at about 28 frames per second.”

This new technology, however, is still in its infancy. Its accuracy has not even exceeded 80%. Further studies and tests are still needed to make it ready for deployment.