Abstract: A learning based super-resolution algorithm for reconstructing face image was proposed. Considering that the similarity of the structures between high resolution (HR) image and corresponding low resolution (LR) image when unfolded on the platform of image library, the input LR image on the built face dictionary for reconstruction was decomposed. Then, the face dictionary of LR images is replaced by corresponding one of HR images with same coefficients. In the coefficients evaluation step, the principal component analysis (PCA) method is used and the total variation (TV) is added as the constraint. The experiment results show that the proposed algorithm could well preserve the faith to the original image and the reconstructed face image is more suitable to be observed by human eyes.