Chapter 11 - On the effect of illumination and face recognition

Abstract

The chapter focuses on the effect of illumination on face recognition. The new concepts and insights introduced in studying illumination modeling previously has benefited in the form of face-recognition algorithms that are robust against illumination variation. The topics discussed include the nonexistence of illumination invariants, Modeling Reflectance and Illumination using Spherical Harmonics, illumination cone and its properties, several recently published algorithms for face-recognition under varying illumination. It is an advantage that human faces do not have more complicated geometry and reflectance. Coupled with the superposition nature of illumination, this allows in utilizing low-dimensional linear appearance models to capture a large portion of image variation due to illumination. The two main elements in modeling illumination effects include the variation in pixel intensity and the formation of shadows. As lighting varies, the radiance at each point on the object's surface also varies according to its reflectance. Linearity makes the algorithms efficient and easy to implement; and the appearance models make the algorithms robust. Even after this many problems are still awaiting formulation and solution. From the face-recognition perspective, there is the problem of detection and alignment, which has been completely ignored. It is difficult to make these processes robust under illumination and a solution to this problem would impact other related research areas such as video face recognition. Because face tracking is an integral part of video face recognition, it is also challenging to develop a tracker that is strong against illumination variation. Pose, expression, occlusion, aging, and other factors must also be considered apart from illumination. Other important issues include photo-realistic simulation of human faces and face recognition using lighting priors. The chapter also provides table and graphs comparing various recognition methods and presenting the analysis using spherical harmonics, respectively.