Image super-resolution via multi-resolution image sequence

Abstract

A novel super-resolution reconstruction algorithm of multi-resolution image sequence integrating the improved super-resolution reconstruction based on neighbor embedding with scale invariant feature transform (SIFT) is proposed in this paper. Firstly, SIFT key points in images are extracted. Then SIFT-feature-based image registration is used to map input high-resolution images to target low-resolution images. Secondly, the mapped images are used as training images and the neighbor embedding is adopted to reconstruct the high-resolution image. The proposed method performs well for problems caused by image deformation, change in viewpoints and change in illumination, which ruin the quality of image super-resolution. Experiments show that the proposed method performs better in terms of lower quantitative errors and better high-frequency information preservation.