Direct kernel PCA with RBF neural networks for face recognition

The conventional kernel PCA does not really nonlinearly maps an input image into a high-dimensional feature space. Rather, it chooses a kernel function a priori and computes the principal components indirectly within the input space spanned by the image pixels. Thus method does not consider the structural information of the input images in the feature space. Therefore, the computed principal components have less discriminating power. In this paper, a new kernel PCA, referred to as the direct kernel PCA (DKPCA), is proposed for face recognition, which explicitly maps an input image nonlinearly into a feature space and then computes the principal components directly in the mapped space. Therefore, this method considers the structural information of the input images in the feature space for computation of principal components, leading to have higher discriminating power. We have designed RBF neural networks for classification of input images based on the computed principal components. The proposed method is evaluated on the ORL database. The results indicate that the proposed method is able to achieve excellent performance and outperforms some of the reported methods using the conventional kernel PCA.