Choosing Multi-illumination training Images based on the degree of linear independency

Multi-illumination training images are usually used in robust face recognition against light variation.Usually a large number of training images are taken under sophisticated imaging system. There definitely exists a large amount of redundancy in these images. Can we pick out several representative samples from them as training images instead of using the whole set? Or even, can we find a better scheme for putting lights when taking or synthesizing these training images? In this paper we propose a method wherewe study the degree of linear independency of face images under different illuminations, and prove that images with different linear independency has different contribution in spanning the illumination subspace.A relatively good combination of images with different linear independency is proposed after analysis and experiments. It also brings forward guidance for light control in obtaining training images.