Boosting Kernel Common Vector Method for Face Recognition

The common vector (CV) method is a linear subspace classifier for datasets, such as those arising in image and word recognition. In this approach, a class subspace is modeled from the common features of all samples in the corresponding class. Since the class subspaces are modeled as a separate subspace in feature domain, there is overlapping between these subspaces and there is loss of information in the common vector of a class. The CVs of a class does not consider its neighboring classes and the overlapping region in between the classes to model the class subspace. This reduces the recognition performance in terms of accuracy. In multi-class problems, within-class and between-class scatter should be considered in classification criterion. But in CV method these scatters are not considered for classification. Generally a linear subspace classifier fails to extract the non-linear features of samples which describe the complexity of face image due to illumination, facial expressions and pose variations. In this paper, we propose a new method called “Boosting kernel common vector method” which solves the above problems by means of its appealing properties. First the discriminatory information obtained from between-class scatter considered the neighboring classes. Second like all kernel methods, it handles non-linearity in a disciplined manner. Third by introducing the boosting parameter into the discriminant criterion increases the recognition performance. Fourth the computation of class subspace as the intersection of the null spaces of that class scatter matrix and the range space of the total scatter matrix yields an optimal subspace for representing a class and it has significant discriminant information. Fifth the classifiers are trained by using hardest samples in the training set increase the classifier performance. Experimental results on YaleB face database demonstrate the promising performance of the proposed methodology.