Subspace classifiers are well-known in pattern recognition, which represent pattern classes by linear subspaces spanned by the class specific basis vectors through simple mathematical operations like SVD. Recently, kernel based subspace methods have been proposed to extend the functionalities by directly applying the Kernel Principal Component Analysis (KPCA). The projection variance in kernel space as applied in these earlier proposed kernel subspace methods, however, is not a trustworthy criteria for class discrimination and they simply fail in many recognition problems as we encountered in biometrics research. We address this issue by proposing a learning kernel subspace classifier which attempts to reconstruct data in input space through the kernel subspace projection. While the pre-image methods aiming at finding an approximate pre-image for each input by minimization of the reconstruction error in kernel space, we emphasize the problem of how to estimate a kernel subspace as a model for a specific class. Using the occluded face recognition as examples, our experimental results demonstrated the efficiency of the proposed method.