A kernel view of manifold analysis for face images

This paper presents a new kernel method to analyse the human face images lying on the low dimensional manifold. Physical variations such as pose and illumination are mapped to the sematic feature space using a kernel matrix and an affine matrix. In this kernel method, the local geometry of the image data is modelled as generative units. The global metric information is also preserved. The kernel formulation enables the manifold to be extended to the out-of-sample data points. This provides a powerful tool for nonlinear dimensional reduction, associative image denoising and image synthesis. Extensive experiments are performed to illustrate the theory.