In this paper, a single image multi-scale super-resolution technique is proposed. The concept under study is the learning procedure between steps of amplification in order to predict the next high scale of resolution. The method integrates two different approaches for the prediction of a high resolution multi-scale scheme, a pure interpolation and a gradient regularization. In the first step a pure interpolation is carried out. It is used a prediction scheme with algebraic reconstruction through different scales to produce the high resolution output. In the last step, the residual blur is reduced by a gradient auto-regularization method. The gradients are adapted by using a weight in a neighbour. Precision of method can be controlled by the parameters of an algebraic reconstruction technique (ART). The proposed model avoids the fast decrease of the output resolution as the amplification factor increases. The proposed system was tested with a dictionary. Results show that the output image quality is improved despite of the increment of the scale factor.