EIGENFACES VS FISHERFACES - RECOGNITION USING CLASS-SPECIFIC LINEAR PROJECTION

We develop a face recognition algorithm which is insensitive to large Variation in lighting direction and facial expression. Taking a patter n classification approach, we consider each pixel in an image as a coo rdinate in a high-dimensional space. We take advantage of the observat ion that the images of a particular face, under varying illumination b ut fixed pose, lie in a 3D linear subspace of the high dimensional ima ge space - if the face is a Lambertian surface without shadowing. Howe ver, since faces are not truly Lambertian surfaces and do indeed produ ce self-shadowing, images will deviate from this linear subspace. Rath er than explicitly modeling this deviation, we linearly project the im age into a subspace in a manner which discounts those regions of the f ace with large deviation. Our projection method is based on Fisher's L inear Discriminant and produces well separated classes in a low-dimens ional subspace, even under severe variation in lighting and facial exp ressions. The Eigenface technique, another method based on linearly pr ojecting the image space to a low dimensional subspace, has similar co mputational requirements. Yet, extensive experimental results demonstr ate that the proposed ''Fisherface'' method has error rates that are t ower than those of the Eigenface technique for tests on the Harvard an d Yale Face Databases.