In order to improve the performance of patch prior based natural image restoration, and effectively remove the noise and blur of images, a restoration framework of expected patch log likelihood (EPLL) using spatially constrained Gaussian mixture model was presented. Based on the spatial distribution information of patches, the priors were defined using the spatially constrained Gaussian mixture statistical characteristics. Image restoration was realized based on the global optimization of image patch restoration. Compared with related works, the proposed method performed better in image denoising and deblurring, and preserved details. The performance of the restoration results was evaluated by the objective indicator. The experimental results show that the proposed method is effective and the visual effect of the image restoration is pleased.