Most of super-resolution reconstruction methods are based on the condition that the point spread function is known or is Gaussian fuzzy kernel. However, the real point spread function of low resolution image, caused by random camera shake, is not Gaussian function. To improve the quality of the reconstructed super-resolution image and make it close to the real scene, a blind image super-resolution method is proposed based on L0 norm sparse representation. First, the point spread function in original images is estimated by using the gradient minimization method based on L0 norm, which is then used for removing the fuzzy effect of image in the process of super resolution reconstruction. At last, the backward propagation algorithm is used to make the reconstruction super-resolution image close to the reality. The experimental results show that: compared with bicubic interpolation and multi-dictionary learning method, the proposed method can get more clear reconstruction effect, and the PSNR and MSSIM are improved. The method has been validated by reconstruction test of real pictures.