Single Image Super Resolution with Neighbor Embedding and In-place Patch Matching

In this paper, we present a novel image super-resolution framework based on neighbor embedding, which belongs to the family of learning-based super-resolution methods. Instead of relying on extrinsic set of training images, image pairs are generated by learning self-similarities from the low-resolution input image itself. Furthermore, to improve the efficiency of image reconstruction, the in-place matching is introduced to the process of similar patches searching. The gradual magnification scheme is adopted to upscale the low-resolution image, and iterative back projection is used to reduce the reconstruction error at each step. Experimental results show that our method achieves satisfactory performance not only on reconstruction quality but also on time efficiency, as compared with other super-resolution methods.