Investigating the spatial support of signal and noise in face recognition

We develop a model for face recognition that describes the image as a sum of signal and noise components. We describe each component as a weighted combination of basis functions. In this paper we investigate the effect of the degree of localization of these basis functions: each might describe the whole image (describe global pixel covariance) or only a small part of the face (describe only local pixel covariance). We find that performance improves when he signal is treated more locally: there is independent information about identity at every position in the image. However, performance decreases when noise is treated more locally: global factors such as pose and illumination conditions can only be understood by looking at a large region of the face. We demonstrate competitive results on several databases using an optimal combination of local signal and global noise models and compare to contemporary approaches.