Neural network based single image super resolution

Abstract

In this paper a novel learning based technique for single image super resolution (SR) is proposed. We model the relationship between available low resolution (LR) image and desired high resolution (HR) image as multi-scale markov random field (MSMRF). We re-formulate the SR problem in terms of learning the mapping between LR-MRF and HR-MRF, which is generally non-linear. Instead of learning MSMRF parameters we use artificial neural networks to learn the desired mapping. The results compare favorably to more complex stat-of-the art techniques for 2 × 2 and 3 × 3 SR problem. We solve the SR problem using optical zoom as a cue by the proposed algorithm as well. The results on experiments with real data are presented.