A spatial neighbourhood based learning setup for super resolution

We formulate the problem of single image super resolution (SR) in terms of learning a single but general nonlinear function. This function takes a low resolution (LR) image patch input and predicts the high resolution (HR) image pixels corresponding to the center pixel of the patch. For training, we use a LR version of an input image, and the given image pixels as target, thus obviating the need for ground truth or explicit search for self similar multiscale patches within a given image. The results compare favorably to more complex state of the art techniques for both noiseless and noisy images. The function needs to be learnt only once using some image, and can also be applied to several other images. We also confirm that spatial Markovian assumption, which is used in methods such as MRF based SR, holds by observing only marginal improvements with increase in LR patch size.