Facial Metrical and Caricature-Pattern-Based Learning in Neural Network System for Face Recognition

Face recognition technology has been an increasingly important module in security systems. A challenging problem is how to extract features tolerant to the appearance variables such as changes in shape, illumination, and occlusion. Extracted metrical features of facial caricatures that are combined with their facial photographs in the training set are examined. The facial caricature is a personal representative amplifying perceptually significant information of individuals. Unlike Eigenfaces, Fisherfaces, and Laplacianfaces, the twenty-nine metrical features that used in this study do not depend upon illumination and occlusion variables. Our results show that facial caricature-trained neural networks outperform significantly of those only facial photograph-trained neural networks.