Many applications have benefited from powerful super- resolution (SR) imaging model, and it is challenging when the missing information in the input low resolution (LR) image. In this paper, we propose a new method to reconstruct a high resolution (HR) image from a low resolution (LR) image based on two dimensional (2D) co-sparse method. The new framework is consisted of three parts. Firstly, divide the nonlinear feature of input LR image into the feature of linear subspaces and learned the LR-HR dictionaries to reduce the artifact. Secondly, 2D co-sparse regularization and self-similarity are developed to strengthen and enhance the image structure. Finally, principal component analysis (PCA) technique is used to reduce the noise in a HR patches. The final SR image can be achieved by reconstructed all HR patches. Simulation results demonstrated a better reconstruction on real image in terms of PSNR and SSIM values, and achieves various improvements compared with other methods.