Face recognition using multispectral random field texture models, color content, and biometric features

Most of the available research on face recognition has been performed using gray scale imagery. This paper presents a novel two-pass face recognition system that uses a multispectral random field texture model, specifically the multispectral simultaneous auto regressive (MSAR) model, and illumination invariant color features. During the first pass, the system detects and segments a face from the background of a color image, and confirms the detection based on a statistically modeled skin pixel map and the elliptical nature of human faces. In the second pass, the face regions are located using the same image segmentation approach on a subspace of the original image, biometric information, and spatial relationships. The determined facial features are then assigned biometric values based on anthropometries, and a set of vectors is created to determine similarity in the facial feature space.