Abstract:  A super-resolution image reconstruction method based on multi-groups of coupled dictionary and alternative learning is presented. In the dictionary learning phase, the image from a training set is viewed as high resolution (HRI). A reduced and re-enlarged version of the HRI is low resolution (LRI). The difference
between HRI and LRI is the residual. The mapping between residual and LRI is obtained from the coupled dictionaries based on the joint data composed of residual patch and LRI patch features. In the reconstruction phase, an enlarged version of the input image is taken as LRI. For each LRI patch, sparse representations and corresponding errors are calculated by using low resolution components of each group of the coupled dictionary. The residual components of coupled dictionary with minimum errors is used to reconstruct the corresponding residual image patch. All reconstructed residual patches together are used to form a residual image, which is then combined with the LRI to produce an HRI. The experimental results demonstrate a satisfied superresolution reconstruction quality.

Key words: super-resolution, sparse representation, multi-dictionary, alternative learning, principal subspace, orthogonal Gaussian mixture model