Face Relighting Based on Multi-spectral Quotient Image and Illumination Tensorfaces

In this paper, a new approach to face relighting by the product of reflectance image and illumination Tensorfaces is proposed. With a pair of multi-spectral images, a near infrared and a visual image, the intrinsic images decomposition can be implemented and corresponding reflectance image is derived. Besides, the illumination images obtained from last step as well as the input visual images constitute a 3-D tensor, on which super-resolution and maximum a posteriori probability estimation are carried out. And then, illumination Tensorfaces under specific light are derived, by which face under target illumination can be synthesized. In contrast to commonly used shape models or shape dependent models, the proposed method only relies on Lambertian assumption and manages to recover reflectance of the face. Besides, compared with the existing methods, i.e. Tensorfaces and Quotient Image, our methods properly preserve the identity of the subject as well as the texture details. Experiments show that the proposed method is not only simple when deriving intrinsic images, but also practical when performing face relighting.