A novel kernel discriminant transformation (KDT) algorithm based on the concept of canonical differences is presented for automatic face recognition applications. For each individual, the face recognition system compiles a multi-view facial image set comprising images with different facial expressions, poses and illumination conditions. Since the multi-view facial images are non-linearly distributed, each image set is mapped into a high-dimensional feature space using a nonlinear mapping function. The corresponding linear subspace, i.e. the kernel subspace, is then constructed via a process of kernel principal component analysis (KPCA). The similarity of two kernel subspaces is assessed by evaluating the canonical difference between them based on the angle between their respective canonical vectors. Utilizing the kernel Fisher discriminant (KFD), a KDT algorithm is derived to establish the correlation between kernel subspaces based on the ratio of the canonical differences of the between-classes to those of the within-classes. The experimental results demonstrate that the proposed classification system outperforms existing subspace comparison schemes and has a promising potential for use in automatic face recognition applications.