3D model-based face recognition system with robustness against illumination changes

In order to make it possible to recognize persons by using face images under any illumination conditions and poses, the authors are developing a new type of a face-recognition system. This is based on a new concept, '2D(query)-3D(DB) scheme.' It acquires an ordinary 2D image as a query, which is then compared with each registered 3D shape and surface reflectance. Two novel technologies are reported here: one is a rangefinder for measuring human faces quickly (2 sec), accurately (0.2mm rms) and densely (pixel by pixel), and the other is an image compensation technique to cope with changes in illumination. The 3D shape enables us to compensate for image changes due to changes in illumination. However, it is difficult to parameterize illumination because the number of parameters, which describe illumination, is infinite due to the number of light, sources. Here the highlight is the authors' reveal that the illumination subspace constructed by quite a few basis images cover almost all the illumination changes. Recognition experiments using images of 42 people under four completely different sets of illumination conditions, for a total of 108 query images, show that our method achieved a first-choice success ratio of 99%.