A Single-Image Super-Resolution Method for Texture Interpolation

In recent years, a number of super-resolution techniques have been proposed. Most of these techniques construct a high resolution image by either combining several low resolution images at sub-pixel misalignments or by learning correspondences between low and high resolution image pairs. In this paper we present a stochastic super-resolution method for color textures from a single image. The proposed algorithm takes advantage of the repetitive nature of textures and the existence of several similar patches within the texture, as well as the color-intensity correlation that often exist in natural images. In the first step of the algorithm the intensity component is interpolated. For each pixel, the missing value is chosen according to a probability distribution constructed from a measure of similarity to other patches in the texture as well as from local features and patch color similarity. In the second stage, the color components are interpolated in a similar manner, using patches of the color channels as well as the already interpolated intensity values. Our conclusion is that the proposed approach outperforms presently available methods.