In the processing of Super-resolution (SR) based on sparse representation, when a testing image is sparsely coded, the coding residual doesn’t simply follow Gaussian or Laplacian distribution. In the maximum likelihood estimation principle, a mixture Gaussian sparse coding model is proposed to solve this problem. Firstly, a weighted l2 norm function is defined to approximate the optimization problem, different weight is defined for different coding residual, and the function is solved by iteratively reweighed sparse coding algorithm. Then the SR model and algorithm are established based on the proposed model, the isomorphic high/low resolution over-complete dictionaries are trained and the sparse representation coefficients of the testing image is learned by the proposed method. At last, the reconstruct method is mended to improve the robustness to noise. The experimental results demonstrate the effectiveness of the proposed method.