Patch Ordering Approach to Single Image Super-resolution Problem

In this paper, we propose a novel patch ordering approach to single image super-resolution (SR) algorithm which is called as patch ordering approach to single image super resolution (POSR). We aimed at selecting more informative high-resolution (HR) and low-resolution (LR) patches for single image SR algorithms based on sparse representation and dictionary learning. Our proposed POSR algorithm, first ordered HR and LR patches for each training images based on minimization of total variation measure (TV). Then, it assigned a sampling step for patch selection in each image. In this way, training patches were extracted based on image texture complexity. This leads to training dictionaries with the high and low resolution more efficiently. Unlike other methods which have used additional restrictions in high resolution image reconstruction phase, the proposed method, has only used the basic assumption of sparse representation super resolution. The experimental results for quantitative criteria (PSNR, RMSE, SSIM and elapsed time), human observation as a qualitative measure and computational complexity verify the improvements offered by the proposed POSR algorithm.