Single image super-resolution by modifying sampling positions

In this paper, a novel single image super-resolution (SR) method is presented using variable sampling positions. The proposed method estimates a sampling position correction vector (SPCV) from the regularly sampled data based on the local gradient of the image. In pursuit of both preserving edge and removing unnatural artifacts in the SR process, non-uniformly sampled data obtained by the SPCV is upscaled using the steering kernel regression algorithm. The proposed SR algorithm restores clear, sharp profile of edge without the reversed gradient or halo effects by modifying the sampling position as well as directionally adaptive interpolation using kernel regression. Because of the unified structure of variable sampling positions and directionally adaptive kernel regression, the propose method does not need to solve any partial difference equations (PDEs) or to use an iterative process, and can easily be adapted to various pre- and post-processing methods for further enhancement.