Local Correlation Classification and Its Application to Face Recognition Across Illumination

In this paper, a real face image is regarded as the result of adding the so-called "standard" face image under an ideal illumination condition to the corresponding "error image", which reflects the imaging difference between the real illumination and the ideal illumination. Furthermore, based on two propositions, we infer that for two images of the same face the correlation between two corresponding areas of the two images will be great enough, while the one between two corresponding areas of two face images of two different individuals will be low. From the viewpoint, a classification algorithm, which is based on a specific definition of correlation between two image areas, is developed. It is computationally tractable and may be regarded as one normalization method. Differing from other normalization methods, this algorithm need not explicitly normalize one face image. The experiment shows that the algorithm is efficient and very excellent for categorizing frontal faces with varying illuminations