To overcome the disadvantages of the unclear results and time consuming in the sparse representation of image super-resolution reconstruction with single redundant dictionary, a single image super-resolution reconstruction method based on multi-dictionary learning and image patches mapping was proposed. In the framework of the traditional sparse representation, firstly the gradient structure information of local image patches was explored, and a large number of training image patches were clustered into several groups by their gradient angles, from those clustered patches the corresponding dictionary pairs were learned. And then the mapping function was computed from low resolution patch to high resolution patch in each clustered group via learned dictionary pairs with the idea of neighbor embedding. Finally the reconstruction process was reduced to a projection of each input patch into the high resolution space by multiplying with the corresponding precomputed mapping function, which improved the images quality with less running time. The experimental results show that the proposed method improves the visual quality significantly, and increases the PSNR (Peak Signal-to-Noise Ratio) at least 0.4 dB compared with the anchored neighborhood regression algorithm.