Low-rank representation for single image superresolution using metric learning

Neighbors embedding is a promising method for single image super-resolution (SR). However, the fixed number of neighbors for different kind of input low resolution (LR) patches may be improper. In addition, the assumption that low resolution space and high resolution (HR) space has similar local geometry leads to improper HR patches are used for reconstruction. In this paper, we propose a novel single image super-resolution method based on low-rank representation and metric learning. Low-rank representation aims to exclude outliers in neighbors, and metric learning aims to learn a linear projection matrix so that LR space with the transformed metric and HR space have similar local structure. Experiments on fourteen images show that our method obtains the best results on most images compared with traditional methods, which illustrates the effectiveness and superiority of the proposed methods.