Example based super-resolution using fuzzy clustering and sparse neighbor embedding

This paper presents a new approach to single-image super resolution, using fuzzy clustering and sparse signal representation. In this method the relationship between low resolution (LR) patches is learnt by fuzzy c-means clustering. By choosing a suitable overcomplete dictionary, LR patch can be represented as a sparse linear combination of the elements from the dictionary. So we are finding a sparse representation for each LR patch and use the coefficient to generate the corresponding high resolution (HR) patch. When an input LR patch is given LR training patches in the selected cluster are sorted based on the decreasing value of membership, which is used for finding the neighbors. Then Robust-SLO algorithm and k/K nearest neighbor selection are used for finding optimal weights and neighboring HR training patches for each LR input patch. The experimental results show that the proposed method gives better results quantitatively and subjectively.