Super-resolution via K-means sparse coding

Abstract

Dictionary learning and sparse representation are efficient methods for single-image super-resolution. We propose a new approach to learn a set of dictionaries and then choose the suitable one for a given test image patch of low resolution. Firstly, the training image patches are clustered into K groups with the information of the test image patches. Secondly, a best basis is learned to model each cluster using sparse prior. Finally, we employ this dictionary to estimate the high resolution patch for the given low resolution patch. This method reduces the complexity of dictionary learning greatly and also makes the representation of patches more compact compared to state-of-the-art methods, which learn a universal dictionary. Experimental results show the effectiveness of our method.