Facial recognition technology may be thwarted by photos' variations in pose, illumination, expression and occlusion. But for the first time, a computer algorithm, developed by researchers from Chinese University of Hong Kong, has beat the facial recognition accuracy of humans.
The researchers Chaochao Lu and Xiaoou Tang used the Labeled Faces in the Wild database – a collection of 13,000 faces of 6,000 public figures from the Internet, each labeled with the person's name – as a benchmark. The database provides a range of face images with variations in pose, lighting, expression, race, ethnicity, age, gender, clothing, hairstyles, and other parameters.
The algorithm, named GaussianFace, scored an accuracy rating of 98.52 percent, beating the human average of 97.53 percent. Prior to GaussianFace, the highest score achieved by technology was 97.25 percent.
According to the Physics arXiv Blog, it works by normalizing each face into a 150x120 pixel image using five image landmarks: the position of both eyes, the nose and the two corners of the mouth. It then divides each image into overlapping patches of 25x25 pixels and describes each patch using a vector, which captures its basic features. Then it compare the images looking for similarities.
GaussianFace was trained by using four databases that contain very different images. One of these is the Multi-PIE database, which consists of face images of 337 subjects from 15 different viewpoints under 19 different conditions of illumination taken in four photo sessions. Another is a database called Life Photos, which contains about 10 images of 400 different people.