Sparse Representation Based Image Super‐resolution Using Large Patches

Abstract

This paper addresses the problem of generating a high-resolution image from a low-resolution image. Many dictionary based methods have been proposed and have achieved great success in super resolution application. Most of these methods use small patches as dictionary atoms, and utilize a unified dictionary pair to conduct reconstruction for each patch, which may limit the super resolution performance. We use large patches instead of small ones to combine a dictionary and to conduct patch reconstruction. Since a large patch contains more meaningful information than a small one, the reconstruction result may have more high frequency details. To guarantee the completeness of the dictionary with large patch, the scale of the dictionary should be large as well. To handle the storage and calculation problems with large dictionaries, we adopt a binary encoding method. This method can preserve local information of patches. For each patch in the low-resolution image, we search its similar patches in the low-resolution dictionary to obtain a sub-dictionary. We compute its sparse representation to get the corresponding high-resolution version. Global reconstruction constraint is enforced to eliminate the discrepancy between the SR result and the ground truth. Experimental results demonstrate that our method outperforms other super resolution methods, especially when the magnification factor is large or the image is blurred by white Gaussian noise.