Improved dictionary learning algorithm with mappings for single image super-resolution

In this paper a coupled dictionary and mapping function learning algorithm is proposed for the task of single image super-resolution. The proposed algorithm consists of dictionary learning part and image reconstruction part. In the dictionary training stage we propose a coupled dictionary and mapping function learning using the K-singular value decomposition. Using this approach we try to enforce the similarity between the high resolution and low resolution sparse coefficients. The idea is to enforce the sparse coefficients of the low resolution image along with its dictionary to well reconstruct the high resolution image using high resolution dictionary. Here we propose to use a mapping function learning which can learn the linear mappings between two resolution levels. For the dictionary update stage the best low rank approximation of singular value decomposition is used. In the image reconstruction stage low resolution images are up-scaled to the high resolution by using the invariance of the sparse coefficients of two resolution levels. First sparse coefficients of low resolution are calculated using the low resolution dictionary and mappings and then they are converted to high resolution by using the high resolution dictionary and mappings. The proposed algorithm is compared with current leading super-resolution algorithms and provides state-of-the-art results.