100% accuracy in automatic face detection.

A problem with the automatic face recognition systems being tested in some airport security screening systems is that none can cope with the kind of image variability encountered in the real world. Jenkins and Burton have used a simple averaging technique to increase the accuracy of an industry standard face-recognition algorithm from 54% to 100%. They averaged the images from 20 different photographs for each of 25 male celebrities who were also in a large public online database of 31,077 photographs of famous faces, comprising an average of nine different photos for each of 3628 celebrities - these images were highly variable in their quality and covered a wide range of lighting conditions, facial expressions, poses, and age. Using the FaceVACS (Cognitec Systems GmbH, Dresden, Germany)industry standard face-recognition system that has been widely adopted, they fed this database their averaged images for each of 25 male celebrities who were also in the online database (excluding photos that were both in their sample and in the database). With the averaged images, the database returned the correct identification 100% of the time. When individual photographs were presented to the database the correct identification was returned only ~50% of the time. From their text:

We demonstrated this improvement with a commercially available algorithm and an online face database over which we had no control. We suggest that image averaging enhances performance by stabilizing the face image. With standard photographs, the match tends to be dominated by aspects of the image that are not diagnostic of identity (e.g., lighting and pose). Averaging together multiple photographs of the same person dilutes these transients while preserving aspects of the image that are consistent across photos. The resulting images capture the visual essence of an individual's face and elevate machine performance to the standard of familiar face recognition in humans. It would be technically straightforward to incorporate an average image into identification documents. Doing so would greatly reduce the incidence of face-recognition errors and raise the prospect of a viable automatic face-recognition infrastructure.


Example photographs of Bill Clinton and their average (right). [Image 1, photo by Marc Nozell (www.flickr.com/photos/marcn/534512066); image 2, photo by Roger Goun (www.flickr.com/photos/sskennel/829574139); image 20, photo by Nelson Pavlosky (www.flickr.com/photos/skyfaller/26752190). All photos were used under a Creative Commons license.] Different pictures of a single face can vary enormously, making automatic recognition difficult. Averaging together multiple photos of the same face stabilizes the image, improving performance dramatically.