Single-image super-resolution based on Markov random field and contourlet transform

摘要

Learning-based methods are well adopted in image super-nresolution. In this paper, we propose a new learning-based approachnusing contourlet transform and Markov random field. The proposednalgorithm employs contourlet transform rather than the conventionalnwavelet to represent image features and takes into account the corre-nlation between adjacent pixels or image patches through the Markovnrandom field (MRF) model. The input low-resolution (LR) image isndecomposed with the contourlet transform and fed to the MRF modelntogether with the contourlet transform coefficients from the low- andnhigh-resolution image pairs in the training set. The unknown high-nfrequency components/coefficients for the input low-resolution im-nage are inferred by a belief propagation algorithm. Finally, the in-nverse contourlet transform converts the LR input and the inferrednhigh-frequency coefficients into the super-resolved image. The ef-nfectiveness of the proposed method is demonstrated with the ex-nperiments on facial, vehicle plate, and real scene images. A betternvisual quality is achieved in terms of peak signal to noise ratio andnthe image structural similarity measurement