We present a novel kernel discriminant transformation (KDT) algorithm based on the concept of canonical differences for automatic face recognition. For each individual, the face recognition system compiles a multi-view facial image set comprising images with different facial expressions, poses and illumination conditions. We derive the KDT algorithm based on the well-known kernel Fisher discriminant (KFD), establishing the correlation between kernel subspaces based on the ratio of the canonical differences of the between-classes to those of the within-classes. The results demonstrate the proposed classification system outperforms existing subspace comparison schemes and has a promising potential for use in automatic face recognition applications.