This paper proposes a wavelet-domain Least-square (LS) based algorithm for image superresolution. Beginning with presenting the edge models popularly accepted in the literature

it is demonstrated in this paper that the edges in different scales are similar to each other in form. This property is called the self-similarity of the multiscale edges. Due to the property

it is possible to predict the three subbands of wavelet coefficients. In order to guarantee the stability and effectiveness of the prediction

the least-square method is adopted. The wavelet coefficients obtained so far are not correct where the multiscale edges are not self-similar. So

the correlation correction method is used to reduce the kind of distortion. Once the wavelet coefficients are obtained

the high resolution image can be reconstructed. Because the algorithm properly preserves the geometrical regularity around the edges

the induced image is of high visual quality. Besides

since only the wavelet coefficients near edges need to be predicted

the algorithm is computationally efficient. Simulations demonstrate the performance of the method.