June 29, 2016 -
A team of researchers from Maryland’s U.S. Army Research Laboratory have developed a new technique exploiting thermal-imaging that potentially could help improve facial-recognition performance that is otherwise hindered by makeup.
Developed by Doctors Nathaniel Short, Alex Yuffa, Gorden Videen, and Shuowen Hu, the new method compares visible, conventional thermal and polarimetric-thermal images of faces before and after the application of face paint.
In their analysis, the researchers found that surface materials have relatively little effect on polarimetric-thermal imagery, demonstrating the effectiveness of polarimetric-thermal imaging for recognizing faces with make-up.
The research team describe their findings in The Optical Society (OSA) journal, Applied Optics.
“Our study has demonstrated polarimetric-thermal imaging can be substantially more robust to face paints, and to a degree cosmetics, for facial recognition than visible imaging,” said lead researcher Nathaniel Short, who is on contract with U.S. Army Research Laboratory as part of the Image Processing Branch. “Our experiments show how face paints and cosmetics degrade the performance of traditional facial-recognition methods and we provide a new approach to mitigating this effect using polarimetric-thermal imaging.”
According to Short, the study marks the first effort to determine the impact of face paints and cosmetic materials on the polarimetric-thermal facial signature.
The researchers believe a polarimetric-thermal-based system could significantly enhance facial-recognition capability for surveillance or security applications, with other advantages in nighttime conditions and in the presence of face paints and cosmetics.
Traditional facial-recognition systems are based on matching clear and well-lit photos captured in the broad light.
Recognizing faces using visible-light imaging depends on capturing the reflected light from the edges of facial features.
This can be difficult when faces are covered with cosmetics as they tend to distort the perceived shape of the face and degrade the face-recognition accuracy of visual imaging due to the different spectral properties of color pigmentation.
In comparison, infrared, thermal signature is naturally emitted from the human face and can be attained passively in low-illumination conditions and even if face paints or cosmetics cover the skin’s surface.
In recent years, thermal imaging has been researched as a new modality for face recognition, particularly in low lighting conditions.
Since conventional thermal imaging only measures thermal intensity or temperature of faces, the thermal facial imagery does not capture the comprehensive shape and texture information, and is also easily affected by the temperature of the skin and the environment.
The researchers have been using polarimetric-thermal imaging, a maturing thermal mode that records the polarization-state information of thermal infrared emission, to collect geometric facial data from thermal imagery. This method could provide several advantages over conventional thermal imaging when matching faces with paints or cosmetics, said Short.
“We found cosmetics and other face paints can significantly degrade facial recognition in the visible imagery, but have relatively little impact on the polarimetric-thermal facial signature,” said Short. “We believe non-visible spectral-imaging techniques such as polarimetric-thermal imaging may facilitate robust face recognition under a number of challenging conditions.”
Despite this promising research, Short emphasizes that the development of the new facial-recognition technique is still in its initial stages and that many challenges still exist.
“One of the major challenges is the limitation of the existing polarimetric-thermal facial database,” said Short. “Large sample pools are needed to develop and train complex machine-learning techniques such as neural networks computer programs that attempt to imitate the human brain to make connections and draw conclusions.”
Another key challenge is in developing algorithms that bridge the large modality gap between visible imaging and polarimetric-thermal imaging for cross-spectrum recognition.