An improved method called Generalized Kernel Fisher Discriminant Analysis was used to enhance face recognition. This method combines two types of discriminant information in "double discriminant subspaces" to improve accuracy. The experimental results on a face database showed that this method is effective for recognizing faces.