Sparse representation based image super-resolution on the KNN based dictionaries

Abstract This paper addresses the problem of single image super-resolution (SR). In recent years, sparse representation based SR methods have been proposed and achieved great success. Traditional sparse representation based SR methods learn a unified high-resolution (HR) and low-resolution (LR) dictionary pair. All LR patches share the same dictionary pair to conduct reconstruction to get their corresponding HR patches. The reconstruction process introduces reconstruction error, which limits the performance of SR results. In this paper, to minimize the reconstruction error, we utilize a K-Nearest-Neighbours (KNN) based dictionary pair for each LR patch instead of the unified dictionary pair to conduct the reconstruction process. The KNN based dictionaries are selected among a huge dictionary which contains hundreds of millions of patch pairs. To speed up the KNN retrieval process, we adopt a binary encoding method which preserves local information for the LR patches, and retrieve the KNN of each LR patch in the Hamming space. Besides, since a large patch contains more structured information than a small patch, we utilize large patches instead of small ones as the atoms of dictionaries, which further improves the SR results. Experimental results demonstrate that our method outperforms the existed state-of-the-art methods, especially when the magnification factor is large or the image is blurred.