Single-Image Super-Resolution Based on Morphological Component Analysis and Coupled Dictionary Learning In Wavelet Domain

High resolution image can provide good performance guarantee for other image processing work, so image super-resolution has always been a hot topic of research. Image super-resolution based on sparse representation and dictionary learning is a very hot technology, these method usually using first-order and second-order derivatives to extract features of image, but this extraction method can not effectively extract high-frequency features such as textures. For this problem, we propose a new single-image super-resolution algorithm that combines concepts from image decomposition theory, morphological component analysis, coupled dictionary learning and wavelet-based dictionary construction. The LR image and texture image layers are used to construct coupled dictionaries. Coupled dictionary learning is used to realize the transformation of the sparse representation coefficients of the two feature spaces through the mapping matrix. This learning scheme relaxes the constraint conditions, enhances the mapping relationship between the LR image and HR texture layers, and improves the reconstruction quality. A set of dictionaries are created in the wavelet domain, where a pair of sub-band dictionaries is designed for the structure and texture image layers in each wavelet sub-band. A total of 3 pairs of dictionaries have been constructed for the structure and texture image components. These dictionaries demonstrate the compactness, directionality, and multi-scale nature of the wavelet transform. Hence, our scheme captures image high-frequency features more effectively. Experimental comparisons show superior super-resolution results of the proposed scheme based on the peak signal-to-noise ratio and the structural similarity index.