Illumination changes will make the performance of face recognition system decreasing dramatically in real world environments. In this paper, we focus on the illumination problem in face recognition, researching in-depth how to use illumination subspace to alleviate the negative effect of illumination changes on face recognition, and the main contributions are summarized as follows: 1. We analysis the concept of illumination normalization and present a novel global illumination subspace-based method to perform illumination normalization in face recognition. The global illumination subspace is constructed using the face images of multiple persons under multiple illumination conditions as training set, and it can overcome the shortcomings of existing methods. 2. We introduce the intrinsic image framework into the illumination research in face recognition. So, achieving illumination-invariant equals to decomposing a face image into the product of an illumination image and a reflectance image using the constraints of faces. In this paper, we describe a method to decompose the shadow free reflectance image from a sequence of images, and discuss its application in face recognition. 3.We present a novel gradient illumination subspace-based method to derive the reflectance image from a single face image. This method apply the independent component analysis on training face images without shadows to construct two gradient illumination subspaces, which then can be used to removing shadow edges from the edges maps of the test image. After a reconstruction process, a shadow free face image can be derived from a single test image.
修改评论