Super-resolution reconstruction plays an important role in reconstructing the image details and improving the visual perception. In the most of the conventional learning-based super-resolution, prior knowledge of the input image itself or natural images database is used to solve the super-resolution problem, so the quality of reconstructed images can be further improved. To reach this goal, the information of the image itself and the natural images database are combined. Firstly, the self-similarities across different image scales can be exploited to construct an image pyramid, and the high resolution image is reconstructed only by the input. After that, we learn a dictionary from natural image patches and reconstruct the initial reconstruction one, which is regarded as the input. In the back processing, non-local similarity and iterative back-projection are exploited to further improve the quality. The experiments show that the proposed algorithmachieves better results than other learning-based algorithms in terms of both visual perception and peak signal-to-noise ratio.