Single image super resolution based on sparse domain selection

Abstract

Single image super-resolution (SR) aims to form a high-resolution (HR) image from an input low-resolution (LR) image. The sparse coding-based example learning methods typically assume that the low-resolution and high-resolution features share the same representation coefficients over their own dictionaries. The assumption is a strong constraint which limits the flexibility to model the complexity of feature space and the mapping among them. To solve the problem, this paper proposes a novel single image super-resolution method utilizing sparse domain selection. In the training phase, the efficient mapping between LR and HR coefficients is established by searching the sparse domain among feature spaces spanned by LR–HR dictionaries. Then this mapping and learning HR dictionary are optimized jointly through minimizing the sparse representation error and sparse domain mapping error. During the reconstruction phase, the learned mapping from the input LR feature is applied to the desired HR feature to achieve accurate and stable SR recovery. Experimental results indicate that the proposed approach is more capable of modeling the relationships from feature of LR to those of HR, thus improves the quality of reconstruction image.