One of the issues with infrared surveillance videos is it can be difficult to recognize the individuals they capture, because a face seen in the infrared light emitted by the body looks different than that same face seen in the visible light reflected off of it. However, researchers at the Karlsruhe Institute of Technology in Germany have developed a method of matching infrared images and visible light images of the same face.
Saquib Sarfraz and Rainer Stiefelhagen developed a deep neural net that they trained to match infrared and visible-light images of faces using a database of 4,585 infrared and visible light images of 82 people created by the University of Notre Dame. The database included images of the participants wearing different expressions and images taken on different days to mirror the ways people's faces change over time and under different circumstances.
Once the deep neural network was trained, it was able to match infrared and visible light faces with 80-percent accuracy, provided it had multiple visible light images against which to compare a thermal image. Accuracy fell to 55 percent if comparisons were one on one. The researchers say improving the system will require assembling a much larger database and building a more powerful network.
From Technology Review
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA