As existing sparse coding methods for single-image easily lead to incorrect geometrical structures in reconstructed images, a sparse coding method combining the incoherence constraint of dictionary and the nonlocal self-similarity constraint of sparse coefficient was proposed . Meanwhile, a double layer reconstruction scheme based a smooth layer (SL) and a texture layer (TL) was presented to overcome the over-smooth edges and blurring problem of the reconstructed images because of introducing the nonlocal self-similarity constraint. The method uses a global non-zero gradient constraint SR model to reconstruct a High Resolution Smooth Image (HRSI), and takes the proposed sparse coding method to recover the HR Texture Image (HRTI). Finally, a global-local constraint optimized model were proposed to improve the quality of the final output image. Experiments indicates that the average values of Peak Signal to Noise(PSNR) and the structural similarity (SSIM) have increased 0.798 7 dB-3.242 4 dB and 0.018 6-0.083 5 as compared with those of some recent representative algorithms。The results demonstrate that the method not only improves the subjective vision obviously, enhances the robustness, but also reconstructs more accurate structures and edges, and receives better reconstruct images.