Facial Recognition using varying lighting from a display.
As phones have become relied upon within society to store sensitive information and facial detection has become more prominent. Having accurate facial detection methods on mobile phones is crucial. Because most phones only have a single camera to use for facial recognition, many facial detection systems can be fooled with a 2D image of a face, instead of the real thing.
This thesis hypothesises that the performance of these signal sensor phones can be improved using a novel approach which builds upon techniques developed in photometric stereo. The approach uses the screen of the mobile device to act as a varying light source, where multiple images can be taken and combined together to obtain information about the surface normal values that make up a face. One major goal for this approach is that it can work on a large variety of devices.
A laptop and mobile application were developed to test this hypothesis, which captures images and performs image processing, then feeds the image to a pre-trained classifier to determine real from fake faces. The app was tested in a range of environments, changing ambient light, number of images captured and movement of the imaged face. Correct classification of fake faces was demonstrated in situations, in which, the current method deployed on phones fails. In other situations, such as high ambient light, the proposed system performed badly.
The proposed method aims to work in conjunction with the existing methods, which are currently in use. Adding another layer of security to to the facial detection systems will incur an increase in processing time but will also provide increased performance and security for these systems.
See the thesis here