Fast Super-Resolution via patchwise Sparse Coding

Single image super-resolution via sparse coding is one of the example-based super-resolution methods. The results are promising. But the relatively low speed hinders its real-time application because of the large and complex computation. We propose a Patchwise Sparse Coding Super-Resolution algorithm to reduce the computation by processing the large mount of patches with patch dimensionality reduction and classification. In our method, the input low-resolution patches are classified and marked by their mean and variance pairs. We use bicubic interpolation, neighbor embedding and sparse coding to different classifications of patches to accelerate the SR process. Additionally we also employ Principle Component Analysis (PCA) dimension reduction to reduce the computation of each patch in the training phase. Finally, the results show us that the computation quantity of the sparse coding image SR is reduced efficiently. The SR time is 17.2s faster than the state-of-the-art SCSR algorithm.