Researchers Develop Face In The Dark

Face recognition is increasingly used, biometric access control to the "tag" photos on social media, but the systems that are actually responsible for this work only during the day and with sufficient light. However, German scientists have now found a solution to face in the dark possible.

Saquib Sarfraz and Rainer Stiefelhagen of the Karlsruhe Institute of Technology have developed a system ( pdf ) developed whereby they analyze dozens of infrared images of a person's face and then compare dozens of images taken in daylight.Researchers feel that the differences in lighting one of the biggest problems that face is facing. Infrared has the advantage that it is insensitive to the present exposure.

In order to compare the images, the scientists have developed a program that uses a "deep neural network system," in order to mimic the operation of a human brain. Through this neural network were over 4500 images which were both made in infrared and daylight are compared within 35 milliseconds to find a match. However, the system also has its problems. As daylight influence can have the photo of someone's face, may infrared images by environmental influences also differ. If it's hot outside or someone have exercised the visions if this person is cool and relaxed.

In the comparison of more than 4500 pictures we worked with 82 people. In 80% of the cases, the match was between the picture in daylight and infrared image correctly. This however requires that there were available several pictures. In case there was only a picture in daylight available, the accuracy decreased to 55%. However, performance can be improved with a larger dataset and powerful network. The conclusion of the study the researchers argue that their approach provides a very practical solution for large-scale surveillance and military applications. They are namely governments, enforcement agencies and the military who may have these datasets and networks.