Fast multi-frame image super-resolution based on MRF

Multi-frame image super-resolution (SR) aims to fuse information in low-resolution (LR) image sequence to compose a high-resolution (HR) one, which is applied in many areas. As an ill-posed problem, SR process suffers from annoying artifacts and remains difficult to balance between the smoothness and the edge preserving. Markov random field (MRF), as a superior method to depict the local characteristics in the image, is applied in the SR process of this paper. This paper firstly proposes an end-to-end SR strategy, which means the corrections in HR space can be derived directly from the reconstruction errors in LR space with no need of projecting into HR space, evidently decreasing the computa-tion amount. Secondly a newly designed MRF-based regu-larization is proposed to realize simultaneous smoothness and edge preserving. The extensive experimental results demon-strate the superior performance of the proposed method.