Abstract: Regularization in the conventional SR process can help to gain a numerical stability and constrain the smoothness of solutions.However,this does not necessarily promise a high quality reconstruction result.This paper proposes a new regularization method for facial image SR,eigen-subspace based regularization.Looking upon the patches of face images as some specific class of signals,their eigen-subspace can be found by Principal Components Analysis,and the super-resolved facial patches are regularized in the orthogonal complement of the eigen-subspace.This eigen-subspace based regularization is combined with the classic reconstruction constraint under a Bayesian framework to produce the high-resolution outcome.Another contribution of this work is we present three iterative algorithms of joint registration and reconstruction estimation,using an affine motion model rather than the most adopted pure translational model.The experimental results illustrate the effectiveness of our approach.