A deformable model for the recognition of human faces under arbitrary illumination

Abstract

Successful recognition systems must be able to handle not only variations in the geometry of the objects they model, but also arbitrary variations in lighting. We propose a deformable template for unoccluded, frontally-viewed human faces that handles both kinds of variation. Lighting is modeled by finding a basis for face space that can be used to synthesize a face image given lighting conditions, or to determine lighting conditions given a face image. Geometric distortions are captured by automatically morphing the input face to the synthesized face. Each of several different face models representing both individuals and average faces was tested on two tasks: discriminating the individual represented by the model from all other faces and nonfaces (about 63 positive examples and 1100 negative examples) and discriminating faces from nonfaces (about 755 positive examples and 450 negative examples). Nonfaces here means patches from random natural and artificial scenes. Each model performed extremely well; for false alarm rates of about 0-3 percent miss rates typically fell in 0-5 percent range indicating that distributions of goodness of fit criteria for negative and positive exemplars are actually very well separated. Nothing about the recognition strategy advocated here is particular to faces; in principle, the model is easily extendible to any other viewpoint or to any other object.