A new method combining MAP and regularization for image super-resolution restoration based on Markov Network is introduced in this paper. The algorithm based on MRF is able to learn a large number of pictures to get an excellent sample from a database. However, it converts the super-resolution restoration problem as a problem of statistical estimation, which computation is relatively large and the running efficiency is poor. Therefore, it is not widely used. The proposed algorithm does not need any statistical assumptions on either the image or the noise. The efficiency of the algorithm is greatly improved by the use of steepest descent method to handle the regularization parameter. Experiments show that the proposed algorithm has a better performance. Compared with a traditional learning-based algorithm, the proposed method has faster operation speed and higher efficiency.