Face recognition in the crowded people under general background

A method for face recognition of acquaintance as a subpattern in a given image is proposed. We consider that the face pattern to be recognized in the input image is approximately an affine transformed (rotated, enlarged and/or reduced, and translated) pattern of a registered original one. In order to estimate the parameters of the affine transformation, the method uses a Karhunen-Loeve (KL) expansion, spatial correlation, and an approximate equation based on Taylor's expansion of affine transformation. In this paper, we deal with two types of pattern representation: ordinary grey level representation and a normalized gradient vector field (NGVF) one. The experimental result shows that our method using NGVF representation is considerably effective.