Non-local neighbor embedding for image super-resolution through FoE features

The super-resolution reconstruction achieved by neighbor embedding usually utilizes manifold learning to map the LR(low resolution) space into the HR(high resolution) space. During the mapping process, the manifold between the LR and HR space is supposed to be similar. Although, the reconstruction method is easy to be implemented, the performance is sensitive to the image features representation and optimal weights. To solve this problem, this paper proposes a novel extended neighbor embedding algorithm for super-resolution reconstruction through FoE(fields of experts) features and non-local self-similar constraint. With the new features, the image edges and textures at various orientations can be responded and the smooth region is ignored. Meanwhile, the optimal weights are calculated with the non-local nearest neighbors. With the non-local self-similar constraint, the optimal weights become more accurate and some undesirable artifacts can be suppressed. Experiments show that the proposed algorithm outperforms other competing algorithms in terms of both quantity and quality.