SISR via trained double sparsity dictionaries

In this paper we present an improved method for single image super-resolution (SISR). The improvement of our method is mainly attributed to the features that we used to train dictionary are wavelets of low resolution (LR) image(s) rather than the first and second derivatives as proposed by Zeyde et al. (2012). As a result, our trained dictionary pair has the property of double sparsity. That means our method can use relatively small training data set to obtain the dictionary with better adaptability to variant natural images. A number of comparison experiments on true images show our method achieves better generalization ability than that proposed in Zeyde et al. (2012).