Improved Kernel Common Vector Method for Face Recognition Varying in Background Conditions

Abstract

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 subspace are modeled as a separate subspace for each class in feature domain, there is overlapping between these subspaces and also loss of information in the common vector of a class. This reduces the recognition performance. In multi-class problems, within-class and between-class scatter should be considered in classification criterion. Since the within class scatter $S_{W}^{}$ and between class scatter $S_{B}^{^{} }$ followed in Discriminative Common Vector method (DCV) are based on the assumption that all classes have similar covariance structures, these class scatters cannot be followed in CV method. 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 “Improved kernel common vector method” which solves the above problems by means of its appealing properties. First the inclusion of boosting parameters in the proposed between-class and within-class scatters consider the neighboring class subspaces and also consider a sample of a class with samples of other classes. This increases the recognition performance. Second the obtained common vector by using the above proposed scatter spaces has more significant discriminative information which also increases the recognition performance. Third like all kernel methods, it handles non-linearity in a disciplined manner which extracts the non-linear features of samples representing the complexity of face images. Experimental results on Yale B face database demonstrate the promising performance of the proposed methodology.