A 2D Hidden Markov Model for Patch-based Super Resolution

[[abstract]]Super resolution is developed to enhance the resolution of images and various kinds of learning based methods were proposed to magnify a single image. This paper presents a 2D hidden Markov model which could do super resolution by using learned image patch pair database. The image patch pairs store the correspondence relation of high-frequency information between low resolution (LR) patches and high resolution (HR) patches. For each input LR patch, the top five similar LR candidate patches in database are searched to construct a 3D cube which can then be modeled by the proposed 2D hidden Markov model (HMM). A novel 2D Viterbi algorithm is developed to find the optimal LR candidate patches that are the most compatible with each other. The resulting super resolution image could be formed by pasting back the corresponding HR patches from patch pair database according to the positions of found optimal LR patches. By objective comparisons of PSNRs/SSIMs and subjective judgment of the generated super resolution images, the proposed 2D HMM method is superior to the traditional interpolation methods and some existing state-of-the-art methods